def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files, field="data")
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = XLNetConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
tokenizer = XLNetTokenizerFast.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = XLNetForQuestionAnswering.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
question_column_name = "question" if "question" in column_names else column_names[0]
context_column_name = "context" if "context" in column_names else column_names[1]
answer_column_name = "answers" if "answers" in column_names else column_names[2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
# Training preprocessing
def prepare_train_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[question_column_name if pad_on_right else context_column_name],
examples[context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=data_args.max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append(
[
0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
for k, s in enumerate(sequence_ids)
]
)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != context_idx:
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != context_idx:
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(token_end_index + 1)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
if training_args.do_train:
train_dataset = datasets["train"].map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Validation preprocessing
def prepare_validation_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[question_column_name if pad_on_right else context_column_name],
examples[context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=data_args.max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append(
[
0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
for k, s in enumerate(sequence_ids)
]
)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
validation_dataset = datasets["validation"].map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
data_collator = default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding(tokenizer)
# Post-processing:
def post_processing_function(examples, features, predictions):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=training_args.output_dir,
is_world_process_zero=trainer.is_world_process_zero(),
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [
{"id": k, "prediction_text": v, "no_answer_probability": scores_diff_json[k]}
for k, v in predictions.items()
]
else:
formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]
references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in datasets["validation"]]
return EvalPrediction(predictions=formatted_predictions, label_ids=references)
# TODO: Once the fix lands in a Datasets release, remove the _local here and the squad_v2_local folder.
current_dir = os.path.sep.join(os.path.join(__file__).split(os.path.sep)[:-1])
metric = load_metric(os.path.join(current_dir, "squad_v2_local") if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions, references=p.label_ids)
# Initialize our Trainer
trainer = QuestionAnsweringTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=validation_dataset if training_args.do_eval else None,
eval_examples=datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets:
train_dataset = datasets.load_dataset("common_voice",
data_args.dataset_config_name,
split=data_args.train_split_name,
cache_dir=model_args.cache_dir)
eval_dataset = datasets.load_dataset("common_voice",
data_args.dataset_config_name,
split="test",
cache_dir=model_args.cache_dir)
# Create and save tokenizer
#chars_to_ignore_regex = f'[{"".join(data_args.chars_to_ignore)}]'
chars_to_ignore_regex = data_args.chars_to_ignore_regex
logger.info('chars_to_ignore_regex = %r', chars_to_ignore_regex)
def remove_special_characters(batch):
batch["text"] = re.sub(chars_to_ignore_regex, "",
batch["sentence"]).lower() + " "
return batch
train_dataset = train_dataset.map(remove_special_characters,
remove_columns=["sentence"])
eval_dataset = eval_dataset.map(remove_special_characters,
remove_columns=["sentence"])
def extract_all_chars(batch):
all_text = " ".join(batch["text"])
vocab = list(set(all_text))
return {"vocab": [vocab], "all_text": [all_text]}
vocab_train = train_dataset.map(
extract_all_chars,
batched=True,
batch_size=-1,
keep_in_memory=True,
remove_columns=train_dataset.column_names,
)
vocab_test = train_dataset.map(
extract_all_chars,
batched=True,
batch_size=-1,
keep_in_memory=True,
remove_columns=eval_dataset.column_names,
)
vocab_list = list(
set(vocab_train["vocab"][0]) | set(vocab_test["vocab"][0]))
vocab_dict = {v: k for k, v in enumerate(vocab_list)}
vocab_dict["|"] = vocab_dict[" "]
del vocab_dict[" "]
vocab_dict["[UNK]"] = len(vocab_dict)
vocab_dict["[PAD]"] = len(vocab_dict)
with open("vocab.json", "w") as vocab_file:
json.dump(vocab_dict, vocab_file)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
tokenizer = Wav2Vec2CTCTokenizer(
"vocab.json",
unk_token="[UNK]",
pad_token="[PAD]",
word_delimiter_token="|",
)
feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1,
sampling_rate=16_000,
padding_value=0.0,
do_normalize=True,
return_attention_mask=True)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor,
tokenizer=tokenizer)
model = Wav2Vec2ForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
activation_dropout=model_args.activation_dropout,
attention_dropout=model_args.attention_dropout,
hidden_dropout=model_args.hidden_dropout,
feat_proj_dropout=model_args.feat_proj_dropout,
mask_time_prob=model_args.mask_time_prob,
gradient_checkpointing=model_args.gradient_checkpointing,
layerdrop=model_args.layerdrop,
ctc_loss_reduction="mean",
pad_token_id=processor.tokenizer.pad_token_id,
vocab_size=len(processor.tokenizer),
)
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(
data_args.max_train_samples))
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# Preprocessing the datasets.
# We need to read the aduio files as arrays and tokenize the targets.
def speech_file_to_array_fn(batch):
speech_array, sampling_rate = torchaudio.load(batch["path"])
batch["speech"] = resampler(speech_array).squeeze().numpy()
batch["sampling_rate"] = 16_000
batch["target_text"] = batch["text"]
return batch
train_dataset = train_dataset.map(
speech_file_to_array_fn,
remove_columns=train_dataset.column_names,
num_proc=data_args.preprocessing_num_workers,
)
eval_dataset = eval_dataset.map(
speech_file_to_array_fn,
remove_columns=eval_dataset.column_names,
num_proc=data_args.preprocessing_num_workers,
)
def prepare_dataset(batch):
# check that all files have the correct sampling rate
assert (
len(set(batch["sampling_rate"])) == 1
), f"Make sure all inputs have the same sampling rate of {processor.feature_extractor.sampling_rate}."
batch["input_values"] = processor(
batch["speech"],
sampling_rate=batch["sampling_rate"][0]).input_values
# Setup the processor for targets
with processor.as_target_processor():
batch["labels"] = processor(batch["target_text"]).input_ids
return batch
train_dataset = train_dataset.map(
prepare_dataset,
remove_columns=train_dataset.column_names,
batch_size=training_args.per_device_train_batch_size,
batched=True,
num_proc=data_args.preprocessing_num_workers,
)
eval_dataset = eval_dataset.map(
prepare_dataset,
remove_columns=eval_dataset.column_names,
batch_size=training_args.per_device_train_batch_size,
batched=True,
num_proc=data_args.preprocessing_num_workers,
)
# Metric
wer_metric = datasets.load_metric("wer")
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids ==
-100] = processor.tokenizer.pad_token_id
pred_str = processor.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
wer = wer_metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
if model_args.freeze_feature_extractor:
model.freeze_feature_extractor()
# Data collator
data_collator = DataCollatorCTCWithPadding(processor=processor,
padding=True)
# Initialize our Trainer
trainer = CTCTrainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=processor.feature_extractor,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
# save the feature_extractor and the tokenizer
if is_main_process(training_args.local_rank):
processor.save_pretrained(training_args.output_dir)
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
return results
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name.",
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--patience",
default="0",
type=str,
required=False,
)
parser.add_argument(
"--regression_threshold",
default=0,
type=float,
required=False,
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help="The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.",
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Run evaluation during training at each logging step.",
)
parser.add_argument(
"--do_lower_case",
action="store_true",
help="Set this flag if you are using an uncased model.",
)
parser.add_argument(
"--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.",
)
parser.add_argument(
"--per_gpu_eval_batch_size",
default=1,
type=int,
help="Batch size per GPU/CPU for evaluation.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
parser.add_argument(
"--save_steps",
type=int,
default=500,
help="Save checkpoint every X updates steps.",
)
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_output_dir",
action="store_true",
help="Overwrite the content of the output directory",
)
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets",
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank",
)
parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
args = parser.parse_args()
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
args.output_dir
)
)
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Set seed
set_seed(args)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
if args.patience != "0" and args.per_gpu_eval_batch_size != 1:
raise ValueError("The eval batch size must be 1 with PABEE inference on.")
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
print("Total Model Parameters:", sum(param.numel() for param in model.parameters()))
output_layers_param_num = sum(param.numel() for param in model.classifiers.parameters())
print("Output Layers Parameters:", output_layers_param_num)
single_output_layer_param_num = sum(param.numel() for param in model.classifiers[0].parameters())
print(
"Added Output Layers Parameters:",
output_layers_param_num - single_output_layer_param_num,
)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
patience_list = [int(x) for x in args.patience.split(",")]
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
print(f"Evaluation for checkpoint {prefix}")
for patience in patience_list:
result = evaluate(args, model, tokenizer, prefix=prefix, patience=patience)
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub
#
# For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
# behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if model_args.model_name_or_path:
model = AutoModelForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
if data_args.line_by_line:
# When using line_by_line, we just tokenize each nonempty line.
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
return tokenizer(
examples["text"],
padding=padding,
truncation=True,
max_length=data_args.max_seq_length,
# We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
# receives the `special_tokens_mask`.
return_special_tokens_mask=True,
)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
else:
# Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
# We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
# efficient when it receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warn(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of
# max_seq_length.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
for k, t in concatenated_examples.items()
}
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
# remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
# might be slower to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
tokenized_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.audio_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.text_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (
f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
)
text_column_name = data_args.text_column_name
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
else:
batch["target_text"] = batch[text_column_name].lower() + " "
return batch
with training_args.main_process_first(desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[text_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
# load config
config = AutoConfig.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
)
# 4. Next, if no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
tokenizer_kwargs = {
"config": config if config.tokenizer_class is not None else None,
"tokenizer_type": config.model_type if config.tokenizer_class is None else None,
"unk_token": unk_token,
"pad_token": pad_token,
"word_delimiter_token": word_delimiter_token,
}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
)
# adapt config
config.update(
{
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": tokenizer.pad_token_id,
"vocab_size": len(tokenizer),
"activation_dropout": model_args.activation_dropout,
}
)
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
num_workers = data_args.preprocessing_num_workers
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["input_length"],
)
# 7. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
return
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
data_collator = DataCollatorCTCWithPadding(processor=processor)
decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
decay_parameters = [name for name in decay_parameters if "bias" not in name]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if n in decay_parameters],
"weight_decay": training_args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if n not in decay_parameters],
"weight_decay": 0.0,
},
]
optimizer = bnb.optim.Adam8bit(
params=optimizer_grouped_parameters,
lr=training_args.learning_rate,
betas=(training_args.adam_beta1, training_args.adam_beta2),
eps=training_args.adam_epsilon,
)
optimizers = (optimizer, None)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
tokenizer=feature_extractor,
optimizers=optimizers,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples
if data_args.max_train_samples is not None
else len(vectorized_datasets["train"])
)
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = (
data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
)
metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = config_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (data_args.train_file.split(".")[-1]
if data_args.train_file is not None else
data_args.validation_file.split(".")[-1])
if extension == "txt":
extension = "text"
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name,
**tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, **tokenizer_kwargs)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger(
"transformers.tokenization_utils_base")
def tokenize_function(examples):
with CaptureLogger(tok_logger) as cl:
output = tokenizer(examples[text_column_name])
# clm input could be much much longer than block_size
if "Token indices sequence length is longer than the" in cl.out:
tok_logger.warning(
"^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
)
return output
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --block_size xxx."
)
block_size = 1024
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {
k: sum(examples[k], [])
for k in examples.keys()
}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k:
[t[i:i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = lm_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
perplexity = math.exp(metrics["eval_loss"])
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.push_to_hub:
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tags": "text-generation"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
trainer.push_to_hub(**kwargs)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if data_args.eval_data_file is None and training_args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
"and load it from here, using --tokenizer_name")
if model_args.model_name_or_path:
model = AutoModelForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if config.model_type in ["bert", "roberta", "distilbert", "camembert"
] and not data_args.mlm:
raise ValueError(
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
"--mlm flag (masked language modeling).")
if data_args.block_size <= 0:
data_args.block_size = tokenizer.model_max_length
# Our input block size will be the max possible for the model
else:
data_args.block_size = min(data_args.block_size,
tokenizer.model_max_length)
# Get datasets
train_dataset = (get_dataset(
data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir)
if training_args.do_train else None)
eval_dataset = (get_dataset(data_args,
tokenizer=tokenizer,
evaluate=True,
cache_dir=model_args.cache_dir)
if training_args.do_eval else None)
if config.model_type == "xlnet":
data_collator = DataCollatorForPermutationLanguageModeling(
tokenizer=tokenizer,
plm_probability=data_args.plm_probability,
max_span_length=data_args.max_span_length,
)
else:
if data_args.mlm and data_args.whole_word_mask:
data_collator = DataCollatorForWholeWordMask(
tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
else:
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=data_args.mlm,
mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(model=model,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path
if model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path) else
None)
trainer.train(model_path=model_path)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_process_zero():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_lm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
check_output_dir(training_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout",
"attention_dropout")
for p in extra_model_params:
if getattr(training_args, p, None):
assert hasattr(
config, p
), f"({config.__class__.__name__}) doesn't have a `{p}` attribute"
setattr(config, p, getattr(training_args, p))
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=".ckpt" in model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
)
# use task specific params
use_task_specific_params(model, data_args.task)
# set num_beams for evaluation
if data_args.eval_beams is None:
data_args.eval_beams = model.config.num_beams
# set decoder_start_token_id for MBart
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, MBartTokenizer):
assert (data_args.tgt_lang is not None and data_args.src_lang
is not None), "mBart requires --tgt_lang and --src_lang"
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.tgt_lang]
if model_args.freeze_embeds:
freeze_embeds(model)
if model_args.freeze_encoder:
freeze_params(model.get_encoder())
assert_all_frozen(model.get_encoder())
dataset_class = Seq2SeqDataset
# Get datasets
train_dataset = (dataset_class(
tokenizer,
type_path="train",
data_dir=data_args.data_dir,
n_obs=data_args.n_train,
max_target_length=data_args.max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_train else None)
eval_dataset = (dataset_class(
tokenizer,
type_path="val",
data_dir=data_args.data_dir,
n_obs=data_args.n_val,
max_target_length=data_args.val_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_eval or
training_args.evaluation_strategy != EvaluationStrategy.NO
else None)
test_dataset = (dataset_class(
tokenizer,
type_path="test",
data_dir=data_args.data_dir,
n_obs=data_args.n_test,
max_target_length=data_args.test_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_predict else None)
# Initialize our Trainer
compute_metrics_fn = (build_compute_metrics_fn(data_args.task, tokenizer)
if training_args.predict_with_generate else None)
trainer = Seq2SeqTrainer(
model=model,
config=config,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=Seq2SeqDataCollator(tokenizer, data_args,
training_args.tpu_num_cores),
compute_metrics=compute_metrics_fn,
data_args=data_args,
)
# Training
if training_args.do_train:
trainer.train(model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_process_zero():
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
eval_results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
if trainer.is_world_process_zero():
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
save_json(
result,
os.path.join(training_args.output_dir, "eval_results.json"))
eval_results.update(result)
if training_args.do_predict:
logging.info("*** Test ***")
test_output = trainer.predict(test_dataset=test_dataset)
test_metrics = {
k.replace("eval", "test"): v
for k, v in test_output.metrics.items()
}
if trainer.is_world_process_zero():
logger.info("***** Test results *****")
for key, value in test_metrics.items():
logger.info(" %s = %s", key, value)
save_json(
test_metrics,
os.path.join(training_args.output_dir, "test_results.json"))
eval_results.update(test_metrics)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_output.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = lmap(str.strip, test_preds)
write_txt_file(
test_preds,
os.path.join(training_args.output_dir,
"test_generations.txt"))
if trainer.is_world_process_zero():
save_json(eval_results, "all_results.json")
return eval_results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
try:
processor = processors[data_args.task_name]()
label_list = processor.get_labels()
num_labels = len(label_list)
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = AutoModelForMultipleChoice.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (MultipleChoiceDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
task=data_args.task_name,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
) if training_args.do_train else None)
eval_dataset = (MultipleChoiceDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
task=data_args.task_name,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
) if training_args.do_eval else None)
def compute_metrics(p: EvalPrediction) -> Dict:
preds = np.argmax(p.predictions, axis=1)
return {"acc": simple_accuracy(preds, p.label_ids)}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train(model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments,
TrainingArguments, MultiLingAdapterArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args, adapter_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args, adapter_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files, field="data")
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=True,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForQuestionAnswering.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Tokenizer check: this script requires a fast tokenizer.
if not isinstance(tokenizer, PreTrainedTokenizerFast):
raise ValueError(
"This example script only works for models that have a fast tokenizer. Checkout the big table of models "
"at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this "
"requirement")
# Setup adapters
if adapter_args.train_adapter:
# new
if data_args.madx2:
# do not add adapter in the last transformer layers
leave_out = [len(model.bert.encoder.layer) - 1]
else:
leave_out = []
# new
# task_name = data_args.dataset_name or "squad"
task_name = "qa"
# optionally load a pretrained language adapter
if adapter_args.load_lang_adapter:
# resolve language adapter config
lang_adapter_config = AdapterConfig.load(
adapter_args.lang_adapter_config,
non_linearity=adapter_args.lang_adapter_non_linearity,
reduction_factor=adapter_args.lang_adapter_reduction_factor,
leave_out=leave_out)
# # load language adapter from Hub
# lang_adapter_name = model.load_adapter(
# adapter_args.load_lang_adapter,
# config=lang_adapter_config,
# load_as=adapter_args.language,
# )
# new
# load language adapter from path in load_lang_adapter
task_mlm_load_as = 'mlm'
lang_adapter_name = model.load_adapter(
adapter_args.load_lang_adapter,
config=lang_adapter_config,
load_as=task_mlm_load_as,
with_head=False)
else:
lang_adapter_name = None
# check if adapter already exists otherwise add it
if task_name not in model.config.adapters:
# # resolve adapter config
# adapter_config = AdapterConfig.load(
# adapter_args.adapter_config,
# non_linearity=adapter_args.adapter_non_linearity,
# reduction_factor=adapter_args.adapter_reduction_factor,
# )
# new
# resolve adapter config with (eventually) the MAD-X 2.0 option
if adapter_args.adapter_config == "pfeiffer":
from transformers.adapters.configuration import PfeifferConfig
adapter_config = PfeifferConfig(
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
leave_out=leave_out)
elif adapter_args.adapter_config == "pfeiffer+inv":
from transformers.adapters.configuration import PfeifferInvConfig
adapter_config = PfeifferInvConfig(
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
leave_out=leave_out)
elif adapter_args.adapter_config == "houlsby":
from transformers.adapters.configuration import HoulsbyConfig
adapter_config = HoulsbyConfig(
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
leave_out=leave_out)
elif adapter_args.adapter_config == "houlsby+inv":
from transformers.adapters.configuration import HoulsbyInvConfig
adapter_config = HoulsbyInvConfig(
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
leave_out=leave_out)
# load adapter from hub if specified
if adapter_args.load_adapter:
model.load_adapter(adapter_args.load_adapter,
config=adapter_config,
load_as=task_name)
else:
model.add_adapter(task_name, config=adapter_config)
# Set the adapters to be used in every forward pass
if lang_adapter_name:
model.active_adapters = Stack(task_mlm_load_as, task_name)
else:
model.set_active_adapters(task_name)
# Freeze all model weights except of those in this adapter
model.train_adapter(task_name)
else:
if adapter_args.load_adapter or adapter_args.load_lang_adapter:
raise ValueError(
"Adapters can only be loaded in adapters training mode."
"Use --train_adapter to enable adapter_training")
# new
# Put only the adapter after the MHA but not after the FF in the last layer
if data_args.houlsby_MHA_lastlayer and adapter_args.train_adapter and not data_args.madx2 and task_name in model.config.adapters and (
adapter_args.adapter_config == "houlsby"
or adapter_args.adapter_config == "houlsby+inv"):
import torch
from torch.nn import ModuleDict
model.bert.encoder.layer[len(model.bert.encoder.layer) -
1].output.adapters = ModuleDict()
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
else:
column_names = datasets["test"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warn(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length" if data_args.pad_to_max_length else False,
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != (1 if pad_on_right
else 0):
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != (1 if pad_on_right else
0):
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
return tokenized_examples
if training_args.do_train:
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
# Create train feature from dataset
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length" if data_args.pad_to_max_length else False,
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
for i in range(len(tokenized_examples["input_ids"])):
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
context_index = 1 if pad_on_right else 0
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_index else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = datasets["validation"]
if data_args.max_val_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(
range(data_args.max_val_samples))
# Validation Feature Creation
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_val_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
if training_args.do_predict:
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_examples = datasets["test"]
if data_args.max_test_samples is not None:
# We will select sample from whole data
test_examples = test_examples.select(
range(data_args.max_test_samples))
# Test Feature Creation
test_dataset = test_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_test_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
data_collator = (default_data_collator if data_args.pad_to_max_length else
DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=8 if training_args.fp16 else None))
# Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions = postprocess_qa_predictions(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
null_score_diff_threshold=data_args.null_score_diff_threshold,
output_dir=training_args.output_dir,
is_world_process_zero=trainer.is_world_process_zero(),
prefix=stage,
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [{
"id": k,
"prediction_text": v,
"no_answer_probability": 0.0
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = load_metric(
"squad_v2" if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
results = metric.compute(predictions=p.predictions,
references=p.label_ids)
# new (change the metric name f1 to eval_f1 in order to avoid a bug at the evaluation time
results['eval_f1'] = results['f1']
results.pop('f1')
return results
# Initialize our Trainer
trainer = QuestionAnsweringTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
eval_examples=eval_examples if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
do_save_full_model=not adapter_args.train_adapter,
do_save_adapters=adapter_args.train_adapter,
callbacks=[
EarlyStoppingCallback(
early_stopping_patience=data_args.early_stopping_patience)
] if training_args.load_best_model_at_end else None,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
return metrics
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
results = trainer.predict(test_dataset, test_examples)
metrics = results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(
test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
def main(json_path):
yangjie_rich_pretrain_unigram_path = '/ai/223/person/lichunyu/pretrain-models/flat/gigaword_chn.all.a2b.uni.ite50.vec'
yangjie_rich_pretrain_bigram_path = '/ai/223/person/lichunyu/pretrain-models/flat/gigaword_chn.all.a2b.bi.ite50.vec'
yangjie_rich_pretrain_word_path = '/ai/223/person/lichunyu/pretrain-models/flat/ctb.50d.vec'
yangjie_rich_pretrain_char_and_word_path = '/ai/223/person/lichunyu/pretrain-models/flat/yangjie_word_char_mix.txt'
# lk_word_path = '/remote-home/xnli/data/pretrain/chinese/sgns.merge.word'
lk_word_path_2 = '/ai/223/person/lichunyu/pretrain-models/flat/sgns.merge.word_2'
load_dataset_seed = 42
logger = logging.getLogger(__name__)
parser = HfArgumentParser((CustomizeArguments, TrainingArguments))
if json_path:
custom_args, training_args = parser.parse_json_file(
json_file=json_path)
elif len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
json_path = os.path.abspath(sys.argv[1])
custom_args, training_args = parser.parse_json_file(
json_file=json_path)
else:
custom_args, training_args = parser.parse_args_into_dataclasses()
_logger = FastNLPLogger(__name__)
_logger.add_file(custom_args.log_file_path)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[
logging.StreamHandler(sys.stdout),
logging.FileHandler(custom_args.log_file_path)
],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
logger.info(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu} "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
set_seed(training_args.seed)
logger.info('Description: {}'.format(custom_args.description))
if json_path:
logger.info('json file path is : {}'.format(json_path))
logger.info('json file args are: \n' + open(json_path, 'r').read())
over_all_dropout = -1
if custom_args.ff_dropout_2 < 0:
custom_args.ff_dropout_2 = custom_args.ff_dropout
if over_all_dropout > 0:
custom_args.embed_dropout = over_all_dropout
custom_args.output_dropout = over_all_dropout
custom_args.pre_dropout = over_all_dropout
custom_args.post_dropout = over_all_dropout
custom_args.ff_dropout = over_all_dropout
custom_args.attn_dropout = over_all_dropout
if custom_args.lattice and custom_args.use_rel_pos:
custom_args.train_clip = True
# now_time = get_peking_time()
if custom_args.test_batch == -1:
custom_args.test_batch = custom_args.batch // 2
device = torch.device('cuda')
refresh_data = False
datasets, vocabs, embeddings = load_ner(
'/ai/223/person/lichunyu/pretrain-models/golden-horse/data',
'/ai/223/person/lichunyu/pretrain-models/flat/gigaword_chn.all.a2b.uni.ite50.vec',
'/ai/223/person/lichunyu/pretrain-models/flat/gigaword_chn.all.a2b.bi.ite50.vec',
_refresh=True,
index_token=False,
# train_clip=custom_args.train_clip,
char_min_freq=custom_args.char_min_freq,
bigram_min_freq=custom_args.bigram_min_freq,
only_train_min_freq=custom_args.only_train_min_freq,
train_path=custom_args.train_path,
dev_path=custom_args.dev_path,
test_path=custom_args.test_path,
placeholder_path=custom_args.placeholder_path,
logger=logger,
with_test_a=custom_args.with_test_a,
label_word2idx=custom_args.label_word2idx)
if custom_args.gaz_dropout < 0:
custom_args.gaz_dropout = custom_args.embed_dropout
custom_args.hidden = custom_args.head_dim * custom_args.head
custom_args.ff = custom_args.hidden * custom_args.ff
if custom_args.lexicon_name == 'lk':
yangjie_rich_pretrain_word_path = lk_word_path_2
w_list = load_yangjie_rich_pretrain_word_list(
yangjie_rich_pretrain_word_path,
_refresh=False,
_cache_fp='cache/{}'.format(custom_args.lexicon_name))
cache_name = os.path.join(
'cache',
(custom_args.dataset + '_lattice' + '_only_train:{}' +
'_trainClip:{}' + '_norm_num:{}' + 'char_min_freq{}' +
'bigram_min_freq{}' + 'word_min_freq{}' + 'only_train_min_freq{}' +
'number_norm{}' + 'lexicon_{}' + 'load_dataset_seed_{}').format(
custom_args.only_lexicon_in_train, custom_args.train_clip,
custom_args.number_normalized, custom_args.char_min_freq,
custom_args.bigram_min_freq, custom_args.word_min_freq,
custom_args.only_train_min_freq, custom_args.number_normalized,
custom_args.lexicon_name, load_dataset_seed))
datasets, vocabs, embeddings = equip_chinese_ner_with_lexicon(
datasets,
vocabs,
embeddings,
w_list,
yangjie_rich_pretrain_word_path,
_refresh=True,
_cache_fp=cache_name,
only_lexicon_in_train=custom_args.only_lexicon_in_train,
word_char_mix_embedding_path=yangjie_rich_pretrain_char_and_word_path,
number_normalized=custom_args.number_normalized,
lattice_min_freq=custom_args.lattice_min_freq,
only_train_min_freq=custom_args.only_train_min_freq,
with_test_a=custom_args.with_test_a)
avg_seq_len = 0
avg_lex_num = 0
avg_seq_lex = 0
train_seq_lex = []
dev_seq_lex = []
test_seq_lex = []
train_seq = []
dev_seq = []
test_seq = []
for k, v in datasets.items():
max_seq_len = 0
max_lex_num = 0
max_seq_lex = 0
max_seq_len_i = -1
for i in range(len(v)):
if max_seq_len < v[i]['seq_len']:
max_seq_len = v[i]['seq_len']
max_seq_len_i = i
# max_seq_len = max(max_seq_len,v[i]['seq_len'])
max_lex_num = max(max_lex_num, v[i]['lex_num'])
max_seq_lex = max(max_seq_lex, v[i]['lex_num'] + v[i]['seq_len'])
avg_seq_len += v[i]['seq_len']
avg_lex_num += v[i]['lex_num']
avg_seq_lex += (v[i]['seq_len'] + v[i]['lex_num'])
if k == 'train':
train_seq_lex.append(v[i]['lex_num'] + v[i]['seq_len'])
train_seq.append(v[i]['seq_len'])
if v[i]['seq_len'] > 200:
logger.info('train里这个句子char长度已经超了200了')
logger.info(''.join(
list(
map(lambda x: vocabs['char'].to_word(x),
v[i]['chars']))))
else:
if v[i]['seq_len'] + v[i]['lex_num'] > 400:
logger.info('train里这个句子char长度没超200，但是总长度超了400')
logger.info(''.join(
list(
map(lambda x: vocabs['char'].to_word(x),
v[i]['chars']))))
if k == 'dev':
dev_seq_lex.append(v[i]['lex_num'] + v[i]['seq_len'])
dev_seq.append(v[i]['seq_len'])
if k == 'test':
test_seq_lex.append(v[i]['lex_num'] + v[i]['seq_len'])
test_seq.append(v[i]['seq_len'])
logger.info('{} 最长的句子是:{}'.format(
k,
list(
map(lambda x: vocabs['char'].to_word(x),
v[max_seq_len_i]['chars']))))
logger.info('{} max_seq_len:{}'.format(k, max_seq_len))
logger.info('{} max_lex_num:{}'.format(k, max_lex_num))
logger.info('{} max_seq_lex:{}'.format(k, max_seq_lex))
max_seq_len = max(*map(lambda x: max(x['seq_len']), datasets.values()))
show_index = 4
logger.info('raw_chars:{}'.format(
list(datasets['train'][show_index]['raw_chars'])))
logger.info('lexicons:{}'.format(
list(datasets['train'][show_index]['lexicons'])))
logger.info('lattice:{}'.format(
list(datasets['train'][show_index]['lattice'])))
logger.info('raw_lattice:{}'.format(
list(
map(lambda x: vocabs['lattice'].to_word(x),
list(datasets['train'][show_index]['lattice'])))))
logger.info('lex_s:{}'.format(list(
datasets['train'][show_index]['lex_s'])))
logger.info('lex_e:{}'.format(list(
datasets['train'][show_index]['lex_e'])))
logger.info('pos_s:{}'.format(list(
datasets['train'][show_index]['pos_s'])))
logger.info('pos_e:{}'.format(list(
datasets['train'][show_index]['pos_e'])))
for k, v in datasets.items():
if custom_args.lattice:
v.set_input('lattice', 'bigrams', 'seq_len', 'target')
v.set_input('lex_num', 'pos_s', 'pos_e')
v.set_target('target', 'seq_len')
v.set_pad_val('lattice', vocabs['lattice'].padding_idx)
else:
v.set_input('chars', 'bigrams', 'seq_len', 'target')
v.set_target('target', 'seq_len')
if custom_args.norm_embed > 0:
logger.info('embedding:{}'.format(
embeddings['char'].embedding.weight.size()))
logger.info('norm embedding')
for k, v in embeddings.items():
norm_static_embedding(v, custom_args.norm_embed)
if custom_args.norm_lattice_embed > 0:
logger.info('embedding:{}'.format(
embeddings['lattice'].embedding.weight.size()))
logger.info('norm lattice embedding')
for k, v in embeddings.items():
norm_static_embedding(v, custom_args.norm_embed)
mode = {}
mode['debug'] = 0
mode['gpumm'] = custom_args.gpumm
# if custom_args.debug or args.gpumm:
# fitlog.debug()
dropout = defaultdict(int)
dropout['embed'] = custom_args.embed_dropout
dropout['gaz'] = custom_args.gaz_dropout
dropout['output'] = custom_args.output_dropout
dropout['pre'] = custom_args.pre_dropout
dropout['post'] = custom_args.post_dropout
dropout['ff'] = custom_args.ff_dropout
dropout['ff_2'] = custom_args.ff_dropout_2
dropout['attn'] = custom_args.attn_dropout
torch.backends.cudnn.benchmark = False
# fitlog.set_rng_seed(args.seed)
torch.backends.cudnn.benchmark = False
# bert_embedding = BertEmbedding(vocabs['lattice'],model_dir_or_name='/root/.fastNLP/embedding/chinese-roberta-wwm-ext-large',requires_grad=True,word_dropout=0.01)
bert_embedding = BertEmbedding(vocabs['lattice'],
model_dir_or_name='cn-wwm',
requires_grad=True,
word_dropout=0.01)
# model = torch.load('/ai/223/person/lichunyu/models/df/ner/flat-2021-08-09-08-40-44-f1_70.pth', map_location=torch.device('cuda'))
model = Lattice_Transformer_SeqLabel(
embeddings['lattice'],
embeddings['bigram'],
custom_args.hidden,
len(vocabs['label']),
custom_args.head,
custom_args.layer,
custom_args.use_abs_pos,
custom_args.use_rel_pos,
custom_args.learn_pos,
custom_args.add_pos,
custom_args.pre,
custom_args.post,
custom_args.ff,
custom_args.scaled,
dropout,
custom_args.use_bigram,
mode,
device,
vocabs,
max_seq_len=max_seq_len,
rel_pos_shared=custom_args.rel_pos_shared,
k_proj=custom_args.k_proj,
q_proj=custom_args.q_proj,
v_proj=custom_args.v_proj,
r_proj=custom_args.r_proj,
self_supervised=custom_args.self_supervised,
attn_ff=custom_args.attn_ff,
pos_norm=custom_args.pos_norm,
ff_activate=custom_args.ff_activate,
abs_pos_fusion_func=custom_args.abs_pos_fusion_func,
embed_dropout_pos=custom_args.embed_dropout_pos,
four_pos_shared=custom_args.four_pos_shared,
four_pos_fusion=custom_args.four_pos_fusion,
four_pos_fusion_shared=custom_args.four_pos_fusion_shared,
bert_embedding=bert_embedding)
with torch.no_grad():
print_info('{}init pram{}'.format('*' * 15, '*' * 15))
for n, p in model.named_parameters():
if 'bert' not in n and 'embedding' not in n and 'pos' not in n and 'pe' not in n \
and 'bias' not in n and 'crf' not in n and p.dim()>1:
try:
if custom_args.init == 'uniform':
nn.init.xavier_uniform_(p)
print_info('xavier uniform init:{}'.format(n))
elif custom_args.init == 'norm':
print_info('xavier norm init:{}'.format(n))
nn.init.xavier_normal_(p)
except:
print_info(n)
exit(1208)
print_info('{}init pram{}'.format('*' * 15, '*' * 15))
encoding_type = 'bio' # bmeso
train_ds = NERDataset(datasets['train'])
train_dataloader = DataLoader(train_ds,
batch_size=64,
collate_fn=collate_func,
shuffle=True)
dev_ds = NERDataset(datasets['dev'])
dev_dataloader = DataLoader(dev_ds, batch_size=32, collate_fn=collate_func)
test_ds = NERDataset(datasets['test'])
test_dataloader = DataLoader(test_ds,
batch_size=8,
collate_fn=collate_func)
bert_embedding_param = list(model.bert_embedding.parameters())
bert_embedding_param_ids = list(map(id, bert_embedding_param))
bigram_embedding_param = list(model.bigram_embed.parameters())
gaz_embedding_param = list(model.lattice_embed.parameters())
embedding_param = bigram_embedding_param
if custom_args.lattice:
gaz_embedding_param = list(model.lattice_embed.parameters())
embedding_param = embedding_param + gaz_embedding_param
embedding_param_ids = list(map(id, embedding_param))
non_embedding_param = list(
filter(
lambda x: id(x) not in embedding_param_ids and id(x) not in
bert_embedding_param_ids, model.parameters()))
param_ = [{
'params': non_embedding_param
}, {
'params': embedding_param,
'lr': custom_args.lr * custom_args.embed_lr_rate
}, {
'params': bert_embedding_param,
'lr': custom_args.bert_lr_rate * custom_args.lr
}]
if custom_args.optim == 'adam':
optimizer = optim.AdamW(param_, lr=custom_args.lr)
print('adam')
elif custom_args.optim == 'sgd':
# optimizer = optim.SGD(model.parameters(),lr=args.lr,momentum=args.momentum,
# weight_decay=args.weight_decay)
optimizer = optim.SGD(param_,
lr=custom_args.lr,
momentum=custom_args.momentum,
weight_decay=0.1)
span_f1_metric = SpanFPreRecMetric(vocabs['label'],
pred='pred',
target='target',
seq_len='seq_len',
encoding_type=encoding_type,
only_gross=False)
# scheduler = LambdaLR(optimizer, lambda ep: 1 / (1 + 0.05*ep) )
epoch = 150
model = nn.DataParallel(model)
model.cuda()
for epoch_n in range(epoch):
logger.info(
'======================epoch {}/{}=============================='.
format(epoch_n, epoch))
model.train()
for n, p in model.named_parameters():
if 'bert_embedding' in n:
if epoch_n >= 10:
p.requires_grad = True
else:
p.requires_grad = False
total_train_loss = 0
for step, batch in enumerate(train_dataloader):
#TODO BERT embedding 前20 epoch 冻结
# chars = batch[0].cuda()
target = batch[1].cuda()
bigrams = batch[2].cuda()
seq_len = batch[3].cuda()
lex_num = batch[4].cuda()
# lex_s = batch[5].cuda()
# lex_e = batch[6].cuda()
lattice = batch[7].cuda()
pos_s = batch[8].cuda()
pos_e = batch[9].cuda()
model.zero_grad()
output = model(lattice, bigrams, seq_len, lex_num, pos_s, pos_e,
target)
loss = output['loss']
total_train_loss += loss.item()
loss.backward()
optimizer.step()
# scheduler.step()
logger.info('train loss: ' +
str(total_train_loss / len(train_dataloader)))
model.eval()
total_eval_loss = 0
for step, batch in enumerate(dev_dataloader):
#TODO BERT embedding 前20 epoch 冻结
# chars = batch[0].cuda()
target = batch[1].cuda()
bigrams = batch[2].cuda()
seq_len = batch[3].cuda()
lex_num = batch[4].cuda()
# lex_s = batch[5].cuda()
# lex_e = batch[6].cuda()
lattice = batch[7].cuda()
pos_s = batch[8].cuda()
pos_e = batch[9].cuda()
with torch.no_grad():
output = model(lattice, bigrams, seq_len, lex_num, pos_s,
pos_e, target)
pred = output['pred']
loss = output['loss']
total_eval_loss += loss.item()
span_f1_metric.evaluate(pred, target, seq_len)
_span_f1 = span_f1_metric.get_metric()
current_ckpt = training_args.output_dir + '/flat-' + datetime.datetime.now(
).strftime("%Y-%m-%d-%H-%M-%S") + '-f1_' + str(int(
_span_f1['f'] * 100)) + '.pth'
logger.info(_span_f1)
logger.info('eval loss: ' + str(total_eval_loss / len(dev_dataloader)))
if epoch_n >= 35:
if custom_args.deploy is True:
logger.info('>>>>>>>>>>>> saving the model <<<<<<<<<<<<<<')
logger.info('model named: {}'.format(current_ckpt))
torch.save(model, current_ckpt)
else:
logger.info(
'>>>>>>>>>>>> saving the state_dict of model <<<<<<<<<<<<<'
)
logger.info('state_dict named: {}'.format(current_ckpt))
torch.save(model.state_dict(), current_ckpt)
logger.info('\n')
print('==============success==============')
pass
def main(json_path=''):
parser = HfArgumentParser((CustomizeArguments, TrainingArguments))
if json_path:
custom_args, training_args = parser.parse_json_file(
json_file=json_path)
elif len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
custom_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
custom_args, training_args = parser.parse_args_into_dataclasses()
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[
logging.StreamHandler(sys.stdout),
logging.FileHandler(custom_args.log_file_path)
],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
logger.info(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu} "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info('Description: {}'.format(custom_args.description))
if json_path:
logger.info('json file path is : {}'.format(json_path))
logger.info('json file args are: \n' + open(json_path, 'r').read())
# last_checkpoint = None
# if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
# last_checkpoint = get_last_checkpoint(training_args.output_dir)
# if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
# raise ValueError(
# f"Output directory ({training_args.output_dir}) already exists and is not empty. "
# "Use --overwrite_output_dir to overcome."
# )
# elif last_checkpoint is not None:
# logger.info(
# f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
# "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
# )
set_seed(training_args.seed)
config = BertConfig.from_pretrained(
custom_args.config_name_or_path
if custom_args.config_name_or_path else custom_args.model_name_or_path,
num_labels=custom_args.num_labels)
model = BertForClassificationByDice.from_pretrained(
custom_args.model_name_or_path, config=config)
tokenizer = BertTokenizer.from_pretrained(
custom_args.tokenizer_name_or_path
if custom_args.tokenizer_name_or_path else
custom_args.model_name_or_path, )
# config = RobertaConfig.from_pretrained(
# custom_args.config_name_or_path if custom_args.config_name_or_path else custom_args.model_name_or_path,
# num_labels=custom_args.num_labels
# )
# model = RobertaForSequenceClassification.from_pretrained(
# custom_args.model_name_or_path,
# config=config
# )
# tokenizer = RobertaTokenizerFast.from_pretrained(
# custom_args.tokenizer_name_or_path if custom_args.tokenizer_name_or_path else custom_args.model_name_or_path,
# )
if custom_args.pickle_data_path:
data = pd.read_pickle(custom_args.pickle_data_path)
df_train = None
df_eval = None
else:
data = None
df_train = pd.read_pickle(custom_args.train_pickle_data_path)
df_eval = pd.read_pickle(custom_args.eval_pickle_data_path)
label_name = custom_args.label_name
text_name = custom_args.text_name
train_dataloader, eval_dataloader = gen_dataloader(
df=data,
df_train=df_train,
df_eval=df_eval,
label_name=label_name,
text_name=text_name,
tokenizer=tokenizer,
per_device_train_batch_size=training_args.per_device_train_batch_size,
per_device_eval_batch_size=training_args.per_device_eval_batch_size,
test_size=custom_args.test_size,
max_length=custom_args.max_length,
)
device = training_args.device if torch.cuda.is_available() else 'cpu'
model = nn.DataParallel(model)
model = model.cuda()
total_bt = time.time()
optimizer = AdamW(model.parameters(), lr=5e-5, eps=1e-6)
total_steps = len(train_dataloader) * training_args.num_train_epochs
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=10,
num_training_steps=total_steps)
# fgm = FGM(model)
for e in range(training_args.num_train_epochs):
logger.info('============= Epoch {:} / {:} =============='.format(
e + 1, training_args.num_train_epochs))
logger.info('Training...')
bt = time.time()
total_train_loss = 0
model.train()
for step, batch in enumerate(train_dataloader):
if step % 50 == 0 and not step == 0:
elapsed = format_time(time.time() - bt)
logger.info(
' Batch {:>5,} of {:>5,}. Elapsed: {:}.'.format(
step, len(train_dataloader), elapsed))
input_ids = batch[0].cuda()
attention_mask = batch[1].cuda()
labels = batch[2].cuda()
model.zero_grad()
output = model(input_ids=input_ids,
attention_mask=attention_mask,
labels=labels)
loss = output.loss
loss = loss.mean()
logits = output.logits
total_train_loss += loss.item()
loss.backward()
# fgm.attack(epsilon=0.9)
# output_adv = model(
# input_ids=input_ids,
# attention_mask=attention_mask,
# labels=labels
# )
# loss_adv = output_adv.loss
# loss_adv.backward()
# fgm.restore()
optimizer.step()
scheduler.step()
avg_train_loss = total_train_loss / len(train_dataloader)
training_time = format_time(time.time() - bt)
logger.info('Average training loss: {0:.2f}'.format(avg_train_loss))
logger.info('Training epcoh took: {:}'.format(training_time))
logger.info('Running Validation...')
bt = time.time()
model.eval()
total_eval_loss = 0
total_eval_f1 = 0
total_eval_acc = 0
total_eval_p = []
total_eval_l = []
for batch in eval_dataloader:
input_ids = batch[0].cuda()
attention_mask = batch[1].cuda()
labels = batch[2].cuda()
with torch.no_grad():
output = model(input_ids=input_ids,
attention_mask=attention_mask,
labels=labels)
loss = output.loss
loss = loss.mean()
logits = output.logits
total_eval_loss += loss.item()
logits = logits.detach().cpu().numpy()
label_ids = labels.to('cpu').numpy()
total_eval_f1 += flat_f1(logits, label_ids)
total_eval_acc += flat_acc(logits, label_ids)
total_eval_p.extend(np.argmax(logits, axis=-1).flatten().tolist())
total_eval_l.extend(label_ids.flatten().tolist())
logger.info(
f'\n{classification_report(total_eval_p, total_eval_l, zero_division=1)}'
)
logger.info(f'Kappa: {cohen_kappa_score(total_eval_l, total_eval_p)}')
avg_val_f1 = total_eval_f1 / len(eval_dataloader)
avg_val_acc = total_eval_acc / len(eval_dataloader)
logger.info('F1: {0:.2f}'.format(avg_val_f1))
logger.info('Acc: {0:.2f}'.format(avg_val_acc))
avg_val_loss = total_eval_loss / len(eval_dataloader)
validation_time = format_time(time.time() - bt)
logger.info('Validation Loss: {0:.2f}'.format(avg_val_loss))
logger.info('Validation took: {:}'.format(validation_time))
current_ckpt = training_args.output_dir + '/bert-' + datetime.datetime.now(
).strftime("%Y-%m-%d-%H-%M-%S") + '-f1_' + str(int(
avg_val_f1 * 100)) + '.pth'
logger.info('Start to save checkpoint named {}'.format(current_ckpt))
if custom_args.deploy is True:
logger.info('>>>>>>>>>>>> saving the model <<<<<<<<<<<<<<')
# torch.save(model.module, current_ckpt)
if isinstance(model, nn.DataParallel):
torch.save(model.module, current_ckpt)
else:
torch.save(model, current_ckpt)
else:
logger.info(
'>>>>>>>>>>>> saving the state_dict of model <<<<<<<<<<<<<')
if isinstance(model, nn.DataParallel):
torch.save(model.module.state_dict(), current_ckpt)
else:
torch.save(model.state_dict(), current_ckpt)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup distant debugging if needed
if data_args.server_ip and data_args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(data_args.server_ip, data_args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# In distributed training, the load_dataset function guarantees that only one local process can concurrently
# download the dataset.
# Downloading and loading xnli dataset from the hub.
if training_args.do_train:
if model_args.train_language is None:
train_dataset = load_dataset("xnli", model_args.language, split="train")
else:
train_dataset = load_dataset("xnli", model_args.train_language, split="train")
label_list = train_dataset.features["label"].names
if training_args.do_eval:
eval_dataset = load_dataset("xnli", model_args.language, split="validation")
label_list = eval_dataset.features["label"].names
if training_args.do_predict:
test_dataset = load_dataset("xnli", model_args.language, split="test")
label_list = test_dataset.features["label"].names
# Labels
num_labels = len(label_list)
# Load pretrained model and tokenizer
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task="xnli",
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
do_lower_case=model_args.do_lower_case,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Preprocessing the datasets
# Padding strategy
if data_args.pad_to_max_length:
padding = "max_length"
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
def preprocess_function(examples):
# Tokenize the texts
return tokenizer(
examples["premise"],
examples["hypothesis"],
padding=padding,
max_length=data_args.max_seq_length,
truncation=True,
)
if training_args.do_train:
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache,
)
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
if training_args.do_eval:
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache,
)
# Get the metric function
metric = load_metric("xnli")
# You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
# predictions and label_ids field) and has to return a dictionary string to float.
def compute_metrics(p: EvalPrediction):
preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
preds = np.argmax(preds, axis=1)
return metric.compute(predictions=preds, references=p.label_ids)
# Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
if data_args.pad_to_max_length:
data_collator = default_data_collator
elif training_args.fp16:
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
else:
data_collator = None
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
checkpoint = None
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
# Check the config from that potential checkpoint has the right number of labels before using it as a
# checkpoint.
if AutoConfig.from_pretrained(model_args.model_name_or_path).num_labels == num_labels:
checkpoint = model_args.model_name_or_path
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.save_model() # Saves the tokenizer too for easy upload
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(eval_dataset=eval_dataset)
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
predictions, labels, metrics = trainer.predict(test_dataset)
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
predictions = np.argmax(predictions, axis=1)
output_test_file = os.path.join(training_args.output_dir, "test_predictions.txt")
if trainer.is_world_process_zero():
with open(output_test_file, "w") as writer:
writer.write("index\tprediction\n")
for index, item in enumerate(predictions):
item = label_list[item]
writer.write(f"{index}\t{item}\n")
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
if data_args.dataset_name == 'ner':
# Downloading and loading a dataset .
datasets = load_dataset('ner_generator.py')
elif data_args.dataset_name == 'pos':
datasets = load_dataset('pos_generator.py')
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
text_column_name = "token"
label_column_name = "tag"
features = datasets["train"].features
label_list = features[label_column_name].names
label2id = {tag: id for id, tag in enumerate(label_list)}
id2label = {id: tag for tag, id in label2id.items()}
num_labels = len(label_list)
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
# def get_label_list(labels):
# unique_labels = set()
# for label in labels:
# unique_labels = unique_labels | set(label)
# label_list = list(unique_labels)
# label_list.sort()
# return label_list
# if isinstance(features[label_column_name].feature, ClassLabel):
# label_list = features[label_column_name].feature.names
# # No need to convert the labels since they are already ints.
# label_to_id = {i: i for i in range(len(label_list))}
# else:
# label_list = get_label_list(datasets["train"][label_column_name])
# label_to_id = {l: i for i, l in enumerate(label_list)}
# num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
label2id=label2id,
id2label=id2label)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=True,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Tokenizer check: this script requires a fast tokenizer.
if not isinstance(tokenizer, PreTrainedTokenizerFast):
raise ValueError(
"This example script only works for models that have a fast tokenizer. Checkout the big table of models "
"at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this "
"requirement")
# Preprocessing the dataset
# Padding strategy
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
padding=padding,
truncation=True,
)
tokenized_inputs["labels"] = [[x] for x in examples[label_column_name]]
return tokenized_inputs
tokenized_datasets = datasets.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = DataCollatorForTokenClassification(
tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
# Metrics
metric = load_metric("seqeval")
def compute_metrics(p):
predictions, labels = p
predictions = np.argmax(predictions, axis=2)
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
true_labels = [[
label_list[l] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
results = metric.compute(predictions=true_predictions,
references=true_labels)
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"]
if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model() # Saves the tokenizer too for easy upload
logger.log(1, metrics)
# trainer.log_metrics("train", metrics)
# trainer.save_metrics("train", metrics)
# trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
logger.log(2, results)
# trainer.log_metrics("eval", results)
# trainer.save_metrics("eval", results)
# Predict
if training_args.do_predict:
logger.info("*** Predict ***")
test_dataset = tokenized_datasets["test"]
predictions, labels, metrics = trainer.predict(test_dataset)
predictions = np.argmax(predictions, axis=2)
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
logger.log(3, metrics)
# trainer.log_metrics("test", metrics)
# trainer.save_metrics("test", metrics)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir,
"test_predictions.txt")
if trainer.is_world_process_zero():
with open(output_test_predictions_file, "w") as writer:
for prediction in true_predictions:
writer.write(" ".join(prediction) + "\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
task_name = data_args.task
lang_id = data_args.language
if task_name is None:
raise ValueError("Set --task should be set to '<xtreme_s_task>' "
"(e.g. 'fleurs-asr', 'mls', 'covost2', 'minds14') ")
if lang_id is None:
raise ValueError(
"Set --language should be set to the language id of the sub dataset "
"config to be used (e.g. 'pl', 'en.tr', 'fr-FR') or 'all'"
" for multi-lingual fine-tuning.")
if data_args.language_group is not None:
if data_args.task != "fleurs-asr":
raise ValueError(
"--language_group should only be used with --task=fleurs-asr")
if data_args.language != "all":
raise ValueError(
"--language_group should only be used with --language=all")
if data_args.target_column_name is None:
target_column_name = TASK_TO_TARGET_COLUMN_NAME[task_name]
else:
target_column_name = data_args.target_column_name
# here we differentiate between tasks with text as the target and classification tasks
is_text_target = target_column_name in ("transcription", "translation")
config_name = ".".join([task_name.split("-")[0], lang_id])
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.audio_column_name not in raw_datasets[
"train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if target_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--target_column_name {target_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--target_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(
range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(
range(data_args.max_eval_samples))
if training_args.do_predict:
raw_datasets["predict"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.predict_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_predict_samples is not None:
raw_datasets["predict"] = raw_datasets["predict"].select(
range(data_args.max_predict_samples))
lang_list = next(iter(raw_datasets.values())).features["lang_id"].names
if not is_text_target:
label_list = next(iter(
raw_datasets.values())).features[target_column_name].names
num_labels = len(label_list)
num_workers = data_args.preprocessing_num_workers
lang_group = data_args.language_group
if lang_group is not None:
with training_args.main_process_first(desc="language group filter"):
lang_group_id = next(iter(
raw_datasets.values())).features["lang_group_id"].str2int(
lang_group)
raw_datasets = raw_datasets.filter(
lambda lang_group: lang_group == lang_group_id,
num_proc=num_workers,
input_columns=["lang_group_id"],
)
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[target_column_name]).lower() + " "
else:
batch["target_text"] = batch[target_column_name].lower() + " "
return batch
if is_text_target:
with training_args.main_process_first(
desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[target_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
if is_text_target:
# 4. (Optional, for ASR and translation) If no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(
vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(
desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
if not config.is_encoder_decoder:
tokenizer_kwargs = {
"config":
config if config.tokenizer_class is not None else None,
"tokenizer_type":
config.model_type
if config.tokenizer_class is None else None,
"unk_token":
unk_token,
"pad_token":
pad_token,
"word_delimiter_token":
word_delimiter_token,
}
else:
tokenizer_kwargs = {}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
if is_text_target:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# adapt config
# (speech translation requires pre-configured seq2seq models)
if task_name != "covost2":
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_zero_infinity": model_args.ctc_zero_infinity,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"activation_dropout": model_args.activation_dropout,
})
if training_args.do_train:
if is_text_target:
config.pad_token_id = tokenizer.pad_token_id
config.vocab_size = len(tokenizer)
else:
label_to_id = {v: i for i, v in enumerate(label_list)}
config.label2id = label_to_id
config.id2label = {
id: label
for label, id in label_to_id.items()
}
config.num_labels = num_labels
# create model
if target_column_name == "transcription":
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
elif config.is_encoder_decoder:
model = AutoModelForSpeechSeq2Seq.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
else:
model = AutoModelForAudioClassification.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[
data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name,
datasets.features.Audio(
sampling_rate=feature_extractor.sampling_rate))
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
if is_text_target:
batch["labels"] = tokenizer(batch["target_text"],
**additional_kwargs).input_ids
else:
batch["labels"] = batch[target_column_name]
batch["lang"] = batch["lang_id"]
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
if training_args.do_train:
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets["train"] = vectorized_datasets["train"].filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["length"],
)
# 7. Next, we can prepare for the training step.
# Let's use the appropriate XTREME-S evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metric = load_metric("xtreme_s", task_name)
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(
f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
)
return
def asr_logits_argmax(logits, labels):
return logits.argmax(dim=-1)
def compute_asr_metric(pred):
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred.predictions)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metric = eval_metric.compute(predictions=pred_str,
references=label_str)
return metric
def compute_classification_metric(pred):
pred_ids = np.argmax(pred.predictions, axis=1)
metric = eval_metric.compute(predictions=pred_ids,
references=pred.label_ids)
return metric
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
if is_text_target:
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
# wait until configs are saved in the main process before loading the processor
if training_args.local_rank != -1:
torch.distributed.barrier()
if is_text_target:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
else:
processor = AutoFeatureExtractor.from_pretrained(
training_args.output_dir)
# Instantiate custom data collator
data_collator = SpeechDataCollatorWithPadding(processor=processor,
pad_labels=is_text_target)
# Initialize Trainer
if target_column_name == "translation":
trainer = Seq2SeqTrainer(
model=model,
data_collator=data_collator,
args=training_args,
preprocess_logits_for_metrics=asr_logits_argmax
if training_args.predict_with_generate else None,
compute_metrics=compute_asr_metric
if training_args.predict_with_generate else None,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
)
else:
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
preprocess_logits_for_metrics=asr_logits_argmax
if is_text_target else None,
compute_metrics=compute_asr_metric
if is_text_target else compute_classification_metric,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(vectorized_datasets["train"]))
metrics["train_samples"] = min(max_train_samples,
len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation on the test set
results = {}
if training_args.do_predict:
logger.info(
f"*** Evaluating on the `{data_args.predict_split_name}` set ***")
if data_args.per_lang_metrics:
# separate the `test` dataset into language-specific subsets and compute metrics for each of them
metrics = {}
average_metrics = defaultdict(list)
for lang_id in range(len(lang_list)):
lang_name = lang_list[lang_id]
with training_args.main_process_first(
desc="per-language dataset filter"):
lang_dataset = vectorized_datasets["predict"].filter(
lambda lang: lang == lang_id,
num_proc=num_workers,
input_columns=["lang"],
)
lang_metrics = trainer.evaluate(lang_dataset)
redundant_metrics = [
"eval_runtime", "eval_samples_per_second",
"eval_steps_per_second", "eval_epoch"
]
for metric_name, value in lang_metrics.items():
average_metrics[metric_name].append(value)
if metric_name not in redundant_metrics:
metrics[f"{metric_name}_{lang_name}"] = value
for metric_name, value in average_metrics.items():
metrics[metric_name] = np.mean(value)
else:
metrics = trainer.evaluate(vectorized_datasets["predict"])
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(vectorized_datasets["predict"]))
metrics["predict_samples"] = min(max_predict_samples,
len(vectorized_datasets["predict"]))
# make sure that the `predict` metrics end up in the log history for the model card
trainer.log(OrderedDict(sorted(metrics.items())))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": task_name,
"tags": [task_name, data_args.dataset_name],
"dataset_args":
f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}, Predict split: {data_args.predict_split_name}",
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
"language": data_args.language,
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
check_output_dir(training_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.parallel_mode == ParallelMode.DISTRIBUTED),
training_args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
set_seed(training_args.seed)
cache_dir = model_args.cache_dir
model_name_or_path = model_args.model_name_or_path
output_dir = training_args.output_dir
do_train = training_args.do_train
do_eval = training_args.do_eval
do_predict = training_args.do_predict
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_name_or_path,
cache_dir=cache_dir,
)
tokenizer.bos_token = tokenizer.cls_token
tokenizer.eos_token = tokenizer.sep_token
train_dataset, eval_dataset = build_datasets(data_args, tokenizer,
cache_dir=cache_dir,
skip_train=not do_train,
skip_eval=not (do_eval or do_predict))
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_name_or_path,
cache_dir=cache_dir,
)
if isinstance(config, EncoderDecoderConfig):
model = AutoModelForSeq2SeqLM.from_pretrained(model_name_or_path, cache_dir=cache_dir)
else:
model = EncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_pretrained_model_name_or_path=model_name_or_path,
decoder_pretrained_model_name_or_path=model_name_or_path,
tie_encoder_decoder=model_args.tie_encoder_decoder,
cache_dir=cache_dir)
update_model_config(model, tokenizer)
if model_args.freeze_embeds:
freeze_embeds(model)
if model_args.freeze_encoder:
freeze_params(model.get_encoder())
assert_all_frozen(model.get_encoder())
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=build_calc_metrics_fn(tokenizer) if do_eval or do_predict else None,
tokenizer=tokenizer,
)
all_metrics = dict()
# Training
if do_train:
logger.info("*** Train ***")
train_result = trainer.train(
model_path=model_name_or_path if os.path.isdir(model_name_or_path) else None
)
metrics = train_result.metrics
metrics["train_n_objs"] = data_args.n_train
trainer.save_model() # this also saves the tokenizer
if trainer.is_world_process_zero():
handle_metrics("train", metrics, output_dir)
all_metrics.update(metrics)
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(output_dir, "trainer_state.json"))
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
tokenizer.save_pretrained(output_dir)
# Evaluation
if do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(
metric_key_prefix="val",
max_length=data_args.val_max_target_length,
num_beams=data_args.eval_beams or model.config.num_beams
)
metrics["val_n_objs"] = data_args.n_val
metrics["val_loss"] = round(metrics["val_loss"], 4)
if trainer.is_world_process_zero():
handle_metrics("val", metrics, output_dir)
all_metrics.update(metrics)
if do_predict:
logger.info("*** Predict ***")
output = trainer.predict(
test_dataset=eval_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.eval_beams or model.config.num_beams
)
metrics = output.metrics
metrics["test_n_objs"] = data_args.n_test
if trainer.is_world_process_zero():
metrics["test_loss"] = round(metrics["test_loss"], 4)
handle_metrics("test", metrics, output_dir)
all_metrics.update(metrics)
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
output.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True
)
predictions = list(map(str.strip, predictions))
write_txt_file(predictions, os.path.join(output_dir, "test_generations.txt"))
if trainer.is_world_process_zero():
save_json(all_metrics, os.path.join(output_dir, "all_results.json"))
return all_metrics
def main(json_path=''):
parser = HfArgumentParser((CustomizeArguments, TrainingArguments))
if json_path:
custom_args, training_args = parser.parse_json_file(json_file=json_path)
elif len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
custom_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
custom_args, training_args = parser.parse_args_into_dataclasses()
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout), logging.FileHandler(custom_args.log_file_path)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
logger.info(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu} "
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info('Description: {}'.format(custom_args.description))
if json_path:
logger.info('json file path is : {}'.format(json_path))
logger.info('json file args are: \n'+open(json_path, 'r').read())
# last_checkpoint = None
# if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
# last_checkpoint = get_last_checkpoint(training_args.output_dir)
# if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
# raise ValueError(
# f"Output directory ({training_args.output_dir}) already exists and is not empty. "
# "Use --overwrite_output_dir to overcome."
# )
# elif last_checkpoint is not None:
# logger.info(
# f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
# "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
# )
set_seed(training_args.seed)
config = BertQueryNerConfig.from_pretrained(
custom_args.config_name_or_path if custom_args.config_name_or_path else custom_args.model_name_or_path,
# num_labels=custom_args.num_labels
)
model = BertQueryNER.from_pretrained(
custom_args.model_name_or_path,
config=config
)
tokenizer = BertTokenizer.from_pretrained(
custom_args.tokenizer_name_or_path if custom_args.tokenizer_name_or_path else custom_args.model_name_or_path,
)
# data = pd.read_pickle(custom_args.pickle_data_path)
# # df_train = pd.read_pickle(custom_args.train_pickle_data_path)
# # df_eval = pd.read_pickle(custom_args.eval_pickle_data_path)
# train_dataloader, eval_dataloader = gen_dataloader(
# df=data,
# # df_train=df_train,
# # df_eval=df_eval,
# tokenizer=tokenizer,
# per_device_train_batch_size=training_args.per_device_train_batch_size,
# per_device_eval_batch_size=training_args.per_device_eval_batch_size,
# test_size=custom_args.test_size,
# max_length=custom_args.max_length,
# )
train_dataloader = get_dataloader('train', 64)
eval_dataloader = get_dataloader('test', 32)
extra_loss = BCEWithLogitsLoss(reduction="none")
extra_dice_loss = MRCDiceLoss(with_logits=True)
# device = training_args.device if torch.cuda.is_available() else 'cpu'
model = nn.DataParallel(model)
model = model.cuda()
total_bt = time.time()
optimizer = AdamW(model.parameters(),
lr = 1e-5,
eps = 1e-8
)
total_steps = len(train_dataloader) * training_args.num_train_epochs
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps = 5,
num_training_steps = total_steps)
weight_sum = custom_args.weight_start + custom_args.weight_end + custom_args.weight_span
weight_start = custom_args.weight_start / weight_sum
weight_end = custom_args.weight_end / weight_sum
weight_span = custom_args.weight_span / weight_sum
# fgm = FGM(model)
for e in range(training_args.num_train_epochs):
logger.info('============= Epoch {:} / {:} =============='.format(e + 1, training_args.num_train_epochs))
logger.info('Training...')
bt = time.time()
total_train_loss = 0
model.train()
for step, batch in enumerate(train_dataloader):
# break
if step % 50 == 0 and not step == 0:
elapsed = format_time(time.time() - bt)
logger.info(' Batch {:>5,} of {:>5,}. Elapsed: {:}. loss: {}'.format(step, len(train_dataloader), elapsed, total_train_loss/step))
input_ids = batch[0].cuda()
token_type_ids = batch[1].cuda()
start_labels = batch[2].cuda()
end_labels = batch[3].cuda()
start_label_mask = batch[4].cuda()
end_label_mask = batch[5].cuda()
match_labels = batch[6].cuda()
# sample_idx = batch[7].cuda()
label_idx = batch[7].cuda()
attention_mask = (input_ids != 0).long()
model.zero_grad()
start_logits, end_logits, span_logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
)
start_loss, end_loss, match_loss = compute_loss(
_loss=extra_dice_loss,
start_logits=start_logits,
end_logits=end_logits,
span_logits=span_logits,
start_labels=start_labels,
end_labels=end_labels,
match_labels=match_labels,
start_label_mask=start_label_mask,
end_label_mask=end_label_mask
)
loss = weight_start * start_loss + weight_end * end_loss + weight_span * match_loss
# loss = output.loss
# logits = output.logits
total_train_loss += loss.item()
loss.backward()
# fgm.attack(epsilon=1.2)
# output_adv = model(
# input_ids=input_ids,
# attention_mask=attention_mask,
# labels=labels
# )
# loss_adv = output_adv.loss
# loss_adv.backward()
# fgm.restore()
optimizer.step()
scheduler.step()
# if step % 50 == 0 and step != 0:
# break
avg_train_loss = total_train_loss / len(train_dataloader)
training_time = format_time(time.time() - bt)
logger.info('Average training loss: {0:.2f}'.format(avg_train_loss))
logger.info('Training epcoh took: {:}'.format(training_time))
logger.info('Running Validation...')
bt = time.time()
model.eval()
total_eval_loss = 0
total_eval_f1 = 0
total_eval_acc = 0
total_eval_p = []
total_eval_l = []
for batch in eval_dataloader:
input_ids = batch[0].cuda()
token_type_ids = batch[1].cuda()
start_labels = batch[2].cuda()
end_labels = batch[3].cuda()
start_label_mask = batch[4].cuda()
end_label_mask = batch[5].cuda()
match_labels = batch[6].cuda()
# sample_idx = batch[7].cuda()
label_idx = batch[7].cuda()
attention_mask = (input_ids != 0).long()
with torch.no_grad():
start_logits, end_logits, span_logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids
)
start_loss, end_loss, match_loss = compute_loss(
_loss=extra_dice_loss,
start_logits=start_logits,
end_logits=end_logits,
span_logits=span_logits,
start_labels=start_labels,
end_labels=end_labels,
match_labels=match_labels,
start_label_mask=start_label_mask,
end_label_mask=end_label_mask
)
loss = weight_start * start_loss + weight_end * end_loss + weight_span * match_loss
total_eval_loss += loss.item()
start_preds, end_preds = start_logits > 0, end_logits > 0
eval_f1 = query_span_f1(start_preds, end_preds, span_logits, start_label_mask, end_label_mask, match_labels)
# logger.info('eval_f1 : {}'.format(eval_f1))
total_eval_f1 += eval_f1
# break
# logger.info(f'\n{classification_report(total_eval_p, total_eval_l, zero_division=1)}')
avg_val_f1 = total_eval_f1 / len(eval_dataloader)
# avg_val_acc = total_eval_acc / len(eval_dataloader)
logger.info('F1: {0:.2f}'.format(avg_val_f1))
# logger.info('Acc: {0:.2f}'.format(avg_val_acc))
avg_val_loss = total_eval_loss / len(eval_dataloader)
validation_time = format_time(time.time() - bt)
logger.info('Validation Loss: {0:.2f}'.format(avg_val_loss))
logger.info('Validation took: {:}'.format(validation_time))
current_ckpt = training_args.output_dir + '/bert-' + datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S") + '-f1_' + str(int(avg_val_f1*100)) + '.pth'
logger.info('Start to save checkpoint named {}'.format(current_ckpt))
if custom_args.deploy is True:
logger.info('>>>>>>>>>>>> saving the model <<<<<<<<<<<<<<')
torch.save(model.module, current_ckpt)
else:
logger.info('>>>>>>>>>>>> saving the state_dict of model <<<<<<<<<<<<<')
torch.save(model.module.state_dict(), current_ckpt)
def main():
# 在 src/transformers/training_args.py中查看所有可能的参数，或将--help标志传递给此脚本。
# 现在，我们保留了不同的参数集，以使关注点更加清晰。
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# 如果我们仅将一个参数传递给脚本，并且它是指向json文件的路径，那么让我们对其进行解析以获取参数。
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"输出目录({training_args.output_dir}) 以及存在，并且不为空"
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# 记录每个进程的日志
logger.warning(
f"使用的 rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu} "
+ f"是否分布式训练: {bool(training_args.local_rank != -1)}, 16-bits 半精度训练: {training_args.fp16}"
)
# 主进程的日志设为verbosity:
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("训练/评估参数 %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub
#
# For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
# behavior (see below)
#
# 在分布式训练中，load_dataset函数可确保只有一个本地进程可以同时下载数据集。
if data_args.dataset_name is not None:
# 从hub下载和加载数据集。
# 首先确定本地缓存了cache文件
cache_script = os.path.join("data", data_args.dataset_name+".py")
if not os.path.exists(cache_script):
raise Exception("请检查本地是否存在相关脚本文件")
datasets = load_dataset(path=cache_script, name=data_args.dataset_config_name, data_dir=data_args.data_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# 加载预训练模型和tokenizer
#
# Distributed training:
# .from_pretrained方法可确保只有一本地个进程可以同时下载模型和vocab。
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("你正从头开始初始化一个新的config.")
# tokenizer的设置
if model_args.tokenizer_name:
if model_args.tokenizer_name == "myroberta":
tokenizer = BertTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
else:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
else:
raise ValueError(
"您正在从头实例化一个新的tokenizer。 此脚本不支持此功能。 "
"您可以用其它形式训练好之后，在这里使用，使用方法: using --tokenizer_name."
)
#模型的设置
if model_args.model_name_or_path:
if model_args.model_name_or_path == 'myroberta':
model = RobertaForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
model = AutoModelForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("从头开始训练一个模型")
model = AutoModelForMaskedLM.from_config(config)
#重设下tokenizer的大小，如果当我们从头训练新模型时，这是必须的
model.resize_token_embeddings(len(tokenizer))
# 处理数据集
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
if data_args.line_by_line:
# 按行处理， tokenize each nonempty line
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# 移除空行
# 收到的数据长度
print(f"收到的数据长度: {[len(t) for t in examples['text']]}")
examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
tokenizer_res = tokenizer(
examples["text"],
padding=padding,
truncation=True,
max_length=data_args.max_seq_length,
# We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
# receives the `special_tokens_mask`.
return_special_tokens_mask=True,
)
print(f"tokenizer之后的数据长度: {print([len(t) for t in tokenizer_res['input_ids']])}")
return tokenizer_res
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
else:
# 否则，我们将tokenize每个文本，然后将它们拼接在一起，然后再将它们分成较小的部分。
# We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
# efficient when it receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
#默认一次处理1000行
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"参数给定的 max_seq_length ({data_args.max_seq_length}) 比模型的 ({tokenizer.model_max_length}) 最大长度长. 使用模型的最大长度 max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# 主要数据处理功能，可拼接数据集中的所有文本并生成max_seq_length的块。
def group_texts(examples):
# 拼接所有文本。
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# 我们删除一小部分，如果模型支持该字段，则可以添加padding，而不是删除，您可以根据需要自定义此部分。
total_length = (total_length // max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
for k, t in concatenated_examples.items()
}
return result
# 注意，使用batched=True`时，此映射一起处理1,000个文本，因此group_texts会丢弃这1,000个文本组中的每一个的余数。 您可以在此处调整该batch_size，但较高的值可能会较慢进行预处理。
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
tokenized_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
# 这部分是随机mask token的设置
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
model_path = (
model_args.model_name_or_path
if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
else None
)
trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** 开始评估 ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help=
"Path to pretrained model or model identifier from huggingface.co/models",
)
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " +
", ".join(glue_processors.keys()),
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written.",
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help=
"Pretrained config name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Pretrained tokenizer name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--cache_dir",
default=None,
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3",
)
parser.add_argument(
"--data_subset",
type=int,
default=-1,
help="If > 0: limit the data to a subset of data_subset instances.")
parser.add_argument("--overwrite_output_dir",
action="store_true",
help="Whether to overwrite data in output directory")
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--dont_normalize_importance_by_layer",
action="store_true",
help="Don't normalize importance score by layers")
parser.add_argument(
"--dont_normalize_global_importance",
action="store_true",
help="Don't normalize all importance scores between 0 and 1",
)
parser.add_argument(
"--try_masking",
action="store_true",
help="Whether to try to mask head until a threshold of accuracy.")
parser.add_argument(
"--masking_threshold",
default=0.9,
type=float,
help=
"masking threshold in term of metrics (stop masking when metric < threshold * original metric value).",
)
parser.add_argument(
"--masking_amount",
default=0.1,
type=float,
help="Amount to heads to masking at each masking step.")
parser.add_argument("--metric_name",
default="acc",
type=str,
help="Metric to use for head masking.")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, sequences shorter padded.",
)
parser.add_argument("--batch_size",
default=1,
type=int,
help="Batch size.")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument("--server_ip",
type=str,
default="",
help="Can be used for distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
# Setup devices and distributed training
if args.local_rank == -1 or args.no_cuda:
args.device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
torch.distributed.init_process_group(
backend="nccl") # Initializes the distributed backend
# Setup logging
logging.basicConfig(
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info("device: {} n_gpu: {}, distributed: {}".format(
args.device, args.n_gpu, bool(args.local_rank != -1)))
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Set seeds
set_seed(args.seed)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name not in glue_processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = glue_processors[args.task_name]()
args.output_mode = glue_output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
output_attentions=True,
cache_dir=args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir,
)
model = AutoModelForSequenceClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir,
)
# Distributed and parallel training
model.to(args.device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Print/save training arguments
os.makedirs(args.output_dir, exist_ok=True)
torch.save(args, os.path.join(args.output_dir, "run_args.bin"))
logger.info("Training/evaluation parameters %s", args)
# Prepare dataset for the GLUE task
eval_dataset = GlueDataset(args, tokenizer=tokenizer, mode="dev")
if args.data_subset > 0:
eval_dataset = Subset(
eval_dataset, list(range(min(args.data_subset,
len(eval_dataset)))))
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.batch_size,
collate_fn=default_data_collator)
# Compute head entropy and importance score
compute_heads_importance(args, model, eval_dataloader)
# Try head masking (set heads to zero until the score goes under a threshole)
# and head pruning (remove masked heads and see the effect on the network)
if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
head_mask = mask_heads(args, model, eval_dataloader)
prune_heads(args, model, eval_dataloader, head_mask)
def main():
# region Argument parsing
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if not (training_args.do_train or training_args.do_eval
or training_args.do_predict):
exit(
"Must specify at least one of --do_train, --do_eval or --do_predict!"
)
# endregion
# region Checkpoints
checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
checkpoint = get_last_checkpoint(training_args.output_dir)
if checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# endregion
# region Logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# endregion
# region Dataset and labels
# Set seed before initializing model.
set_seed(training_args.seed)
# Downloading and loading a dataset from the hub. In distributed training, the load_dataset function guarantee
# that only one local process can concurrently download the dataset.
datasets = load_dataset(
"glue",
data_args.task_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
is_regression = data_args.task_name == "stsb"
if not is_regression:
label_list = datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
if data_args.predict_file is not None:
logger.info("Preparing user-supplied file for predictions...")
data_files = {"data": data_args.predict_file}
for key in data_files.keys():
logger.info(f"Loading a local file for {key}: {data_files[key]}")
if data_args.predict_file.endswith(".csv"):
# Loading a dataset from local csv files
user_dataset = load_dataset("csv",
data_files=data_files,
cache_dir=model_args.cache_dir)
else:
# Loading a dataset from local json files
user_dataset = load_dataset("json",
data_files=data_files,
cache_dir=model_args.cache_dir)
needed_keys = task_to_keys[data_args.task_name]
for key in needed_keys:
assert key in user_dataset[
"data"].features, f"Your supplied predict_file is missing the {key} key!"
datasets["user_data"] = user_dataset["data"]
# endregion
# region Load model config and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# endregion
# region Dataset preprocessing
sentence1_key, sentence2_key = task_to_keys[data_args.task_name]
non_label_column_names = [
name for name in datasets["train"].column_names if name != "label"
]
# Padding strategy
if data_args.pad_to_max_length:
padding = "max_length"
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if config.label2id != PretrainedConfig(
num_labels=num_labels).label2id and not is_regression:
# Some have all caps in their config, some don't.
label_name_to_id = {k.lower(): v for k, v in config.label2id.items()}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
label_to_id = {
i: int(label_name_to_id[label_list[i]])
for i in range(num_labels)
}
else:
logger.warning(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
label_to_id = {label: i for i, label in enumerate(label_list)}
if label_to_id is not None:
config.label2id = label_to_id
config.id2label = {id: label for label, id in config.label2id.items()}
elif data_args.task_name is not None and not is_regression:
config.label2id = {l: i for i, l in enumerate(label_list)}
config.id2label = {id: label for label, id in config.label2id.items()}
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_function(examples):
# Tokenize the texts
args = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*args,
padding=padding,
max_length=max_seq_length,
truncation=True)
return result
datasets = datasets.map(preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache)
if data_args.pad_to_max_length:
data_collator = DefaultDataCollator(return_tensors="tf")
else:
data_collator = DataCollatorWithPadding(tokenizer, return_tensors="tf")
# endregion
# region Metric function
metric = load_metric("glue", data_args.task_name)
def compute_metrics(preds, label_ids):
preds = preds["logits"]
preds = np.squeeze(preds) if is_regression else np.argmax(preds,
axis=1)
result = metric.compute(predictions=preds, references=label_ids)
if len(result) > 1:
result["combined_score"] = np.mean(list(result.values())).item()
return result
# endregion
with training_args.strategy.scope():
# region Load pretrained model
if checkpoint is None:
model_path = model_args.model_name_or_path
else:
model_path = checkpoint
model = TFAutoModelForSequenceClassification.from_pretrained(
model_path,
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# endregion
# region Optimizer, loss and compilation
optimizer = tf.keras.optimizers.Adam(
learning_rate=training_args.learning_rate,
beta_1=training_args.adam_beta1,
beta_2=training_args.adam_beta2,
epsilon=training_args.adam_epsilon,
clipnorm=training_args.max_grad_norm,
)
if is_regression:
loss_fn = tf.keras.losses.MeanSquaredError()
metrics = []
else:
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True)
metrics = ["accuracy"]
model.compile(optimizer=optimizer, loss=loss_fn, metrics=metrics)
# endregion
# region Convert data to a tf.data.Dataset
tf_data = dict()
max_samples = {
"train": data_args.max_train_samples,
"validation": data_args.max_eval_samples,
"validation_matched": data_args.max_eval_samples,
"validation_mismatched": data_args.max_eval_samples,
"test": data_args.max_predict_samples,
"test_matched": data_args.max_predict_samples,
"test_mismatched": data_args.max_predict_samples,
"user_data": None,
}
for key in datasets.keys():
if key == "train" or key.startswith("validation"):
assert "label" in datasets[
key].features, f"Missing labels from {key} data!"
if key == "train":
shuffle = True
batch_size = training_args.per_device_train_batch_size
drop_remainder = True # Saves us worrying about scaling gradients for the last batch
else:
shuffle = False
batch_size = training_args.per_device_eval_batch_size
drop_remainder = False
samples_limit = max_samples[key]
dataset = datasets[key]
if samples_limit is not None:
dataset = dataset.select(range(samples_limit))
data = dataset.to_tf_dataset(
columns=[
col for col in dataset.column_names
if col not in set(non_label_column_names + ["label"])
],
shuffle=shuffle,
batch_size=batch_size,
collate_fn=data_collator,
drop_remainder=drop_remainder,
# `label_cols` is needed for user-defined losses, such as in this example
label_cols="label"
if "label" in dataset.column_names else None,
)
tf_data[key] = data
# endregion
# region Training and validation
if training_args.do_train:
callbacks = [
SavePretrainedCallback(output_dir=training_args.output_dir)
]
if training_args.do_eval and not data_args.task_name == "mnli":
# Do both evaluation and training in the Keras fit loop, unless the task is MNLI
# because MNLI has two validation sets
validation_data = tf_data["validation"]
else:
validation_data = None
model.fit(
tf_data["train"],
validation_data=validation_data,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks,
)
# endregion
# region Evaluation
if training_args.do_eval:
# We normally do validation as part of the Keras fit loop, but we run it independently
# if there was no fit() step (because we didn't train the model) or if the task is MNLI,
# because MNLI has a separate validation-mismatched validation set
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
if data_args.task_name == "mnli":
tasks = ["mnli", "mnli-mm"]
tf_datasets = [
tf_data["validation_matched"],
tf_data["validation_mismatched"]
]
raw_datasets = [
datasets["validation_matched"],
datasets["validation_mismatched"]
]
else:
tasks = [data_args.task_name]
tf_datasets = [tf_data["validation"]]
raw_datasets = [datasets["validation"]]
for raw_dataset, tf_dataset, task in zip(raw_datasets, tf_datasets,
tasks):
eval_predictions = model.predict(tf_dataset)
eval_metrics = compute_metrics(eval_predictions,
raw_dataset["label"])
print(f"Evaluation metrics ({task}):")
print(eval_metrics)
# endregion
# region Prediction
if training_args.do_predict or data_args.predict_file:
logger.info("*** Predict ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = []
tf_datasets = []
raw_datasets = []
if training_args.do_predict:
if data_args.task_name == "mnli":
tasks.extend(["mnli", "mnli-mm"])
tf_datasets.extend(
[tf_data["test_matched"], tf_data["test_mismatched"]])
raw_datasets.extend([
datasets["test_matched"], datasets["test_mismatched"]
])
else:
tasks.append(data_args.task_name)
tf_datasets.append(tf_data["test"])
raw_datasets.append(datasets["test"])
if data_args.predict_file:
tasks.append("user_data")
tf_datasets.append(tf_data["user_data"])
raw_datasets.append(datasets["user_data"])
for raw_dataset, tf_dataset, task in zip(raw_datasets, tf_datasets,
tasks):
test_predictions = model.predict(tf_dataset)
if "label" in raw_dataset:
test_metrics = compute_metrics(test_predictions,
raw_dataset["label"])
print(f"Test metrics ({task}):")
print(test_metrics)
if is_regression:
predictions_to_write = np.squeeze(
test_predictions["logits"])
else:
predictions_to_write = np.argmax(
test_predictions["logits"], axis=1)
output_predict_file = os.path.join(
training_args.output_dir, f"predict_results_{task}.txt")
with open(output_predict_file, "w") as writer:
logger.info(
f"***** Writing prediction results for {task} *****")
writer.write("index\tprediction\n")
for index, item in enumerate(predictions_to_write):
if is_regression:
writer.write(f"{index}\t{item:3.3f}\n")
else:
item = model.config.id2label[item]
writer.write(f"{index}\t{item}\n")
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " + ", ".join(MODEL_TYPES),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models",
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints and predictions will be written.",
)
# Other parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
help="The input data dir. Should contain the .json files for the task."
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--train_file",
default=None,
type=str,
help="The input training file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--predict_file",
default=None,
type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from s3",
)
parser.add_argument(
"--version_2_with_negative",
action="store_true",
help="If true, the SQuAD examples contain some that do not have an answer.",
)
parser.add_argument(
"--null_score_diff_threshold",
type=float,
default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.",
)
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded.",
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.",
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help="The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.",
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument(
"--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
)
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
)
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.",
)
parser.add_argument(
"--verbose_logging",
action="store_true",
help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.",
)
parser.add_argument(
"--lang_id",
default=0,
type=int,
help="language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)",
)
parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
parser.add_argument("--save_steps", type=int, default=500, help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
parser.add_argument(
"--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
args = parser.parse_args()
if args.doc_stride >= args.max_seq_length - args.max_query_length:
logger.warning(
"WARNING - You've set a doc stride which may be superior to the document length in some "
"examples. This could result in errors when building features from the examples. Please reduce the doc "
"stride or increase the maximum length to ensure the features are correctly built."
)
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
args.output_dir
)
)
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
# Make sure only the first process in distributed training will download model & vocab
torch.distributed.barrier()
args.model_type = args.model_type.lower()
config = AutoConfig.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
use_fast=False, # SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
)
model = AutoModelForQuestionAnswering.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
# Make sure only the first process in distributed training will download model & vocab
torch.distributed.barrier()
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
# Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
# remove the need for this code, but it is still valid.
if args.fp16:
try:
import apex
apex.amp.register_half_function(torch, "einsum")
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Save the trained model and the tokenizer
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(model, "module") else model
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = AutoModelForQuestionAnswering.from_pretrained(args.output_dir) # , force_download=True)
# SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
# So we use use_fast=False here for now until Fast-tokenizer-compatible-examples are out
tokenizer = AutoTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case, use_fast=False)
model.to(args.device)
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
if args.do_train:
logger.info("Loading checkpoints saved during training for evaluation")
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c)
for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
else:
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
checkpoints = [args.model_name_or_path]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
# Reload the model
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
model = AutoModelForQuestionAnswering.from_pretrained(checkpoint) # , force_download=True)
model.to(args.device)
# Evaluate
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
results.update(result)
logger.info("Results: {}".format(results))
return results
def main():
FASTA_DATASET = False
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub
#
# For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
# behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "fasta":
FASTA_DATASET = True
datasets = load_dataset_fasta_protbert(data_files,
data_args.max_seq_length)
else:
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
max_length=data_args.max_seq_length)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if model_args.model_name_or_path:
model = AutoModelForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
tokenized_datasets = dict()
for dataset_key, dataset in datasets.items():
# Tokenize
encodings = tokenizer(
dataset['sequences'],
truncation=True,
padding='max_length', # TODO get from args passed in
max_length=data_args.max_seq_length,
return_special_tokens_mask=True,
return_token_type_ids=False,
return_attention_mask=False)
torch_dataset = FastaDataset(encodings)
tokenized_datasets[dataset_key] = torch_dataset
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"]
if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path if
(model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path)) else
None)
trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_mlm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets:
train_dataset = datasets.load_dataset(CORPORA_PATH,
split="train",
cache_dir=model_args.cache_dir)
eval_dataset = datasets.load_dataset(CORPORA_PATH,
split="test",
cache_dir=model_args.cache_dir)
feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1,
sampling_rate=16_000,
padding_value=0.0,
do_normalize=True,
return_attention_mask=True)
processor = CustomWav2Vec2Processor(feature_extractor=feature_extractor)
model = Wav2VecClassifierModel.from_pretrained(
"facebook/wav2vec2-large-xlsr-53",
attention_dropout=0.01,
hidden_dropout=0.01,
feat_proj_dropout=0.0,
mask_time_prob=0.05,
layerdrop=0.01,
gradient_checkpointing=True,
)
if model_args.freeze_feature_extractor:
model.freeze_feature_extractor()
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(
data_args.max_train_samples))
if data_args.max_val_samples is not None:
max_val_samples = min(data_args.max_val_samples, len(eval_dataset))
eval_dataset = eval_dataset.select(range(max_val_samples))
# Preprocessing the datasets.
# We need to read the aduio files as arrays and tokenize the targets.
def speech_file_to_array_fn(batch):
start = 0
stop = SECONDS_STOP
srate = S_RATE
speech_array, sampling_rate = torchaudio.load(batch["file"])
speech_array = speech_array[0].numpy()[:stop * sampling_rate]
batch["speech"] = librosa.resample(np.asarray(speech_array),
sampling_rate, srate)
batch["sampling_rate"] = srate
batch["parent"] = batch["label"]
return batch
train_dataset = train_dataset.map(
speech_file_to_array_fn,
remove_columns=train_dataset.column_names,
num_proc=data_args.preprocessing_num_workers,
)
eval_dataset = eval_dataset.map(
speech_file_to_array_fn,
remove_columns=eval_dataset.column_names,
num_proc=data_args.preprocessing_num_workers,
)
def prepare_dataset(batch):
# check that all files have the correct sampling rate
assert (
len(set(batch["sampling_rate"])) == 1
), f"Make sure all inputs have the same sampling rate of {processor.feature_extractor.sampling_rate}."
batch["input_values"] = processor(
batch["speech"],
sampling_rate=batch["sampling_rate"][0]).input_values
batch["labels"] = batch["parent"]
return batch
train_dataset = train_dataset.map(
prepare_dataset,
remove_columns=train_dataset.column_names,
batch_size=training_args.per_device_train_batch_size,
batched=True,
num_proc=data_args.preprocessing_num_workers,
)
eval_dataset = eval_dataset.map(
prepare_dataset,
remove_columns=eval_dataset.column_names,
batch_size=training_args.per_device_train_batch_size,
batched=True,
num_proc=data_args.preprocessing_num_workers,
)
from sklearn.metrics import classification_report, confusion_matrix
def compute_metrics(pred):
label_idx = LABEL_IDX
label_names = LABEL_NAMES
labels = pred.label_ids.argmax(-1)
preds = pred.predictions.argmax(-1)
acc = accuracy_score(labels, preds)
f1 = f1_score(labels, preds, average='macro')
report = classification_report(y_true=labels,
y_pred=preds,
labels=label_idx,
target_names=label_names)
matrix = confusion_matrix(y_true=labels, y_pred=preds)
print(report)
print(matrix)
wandb.log({
"conf_mat":
wandb.plot.confusion_matrix(probs=None,
y_true=labels,
preds=preds,
class_names=label_names)
})
wandb.log({
"precision_recall":
wandb.plot.pr_curve(y_true=labels,
y_probas=pred.predictions,
labels=label_names)
})
return {"accuracy": acc, "f1_score": f1}
wandb.init(name=training_args.output_dir, config=training_args)
# Data collator
data_collator = DataCollatorCTCWithPadding(processor=processor,
padding=True)
# Initialize our Trainer
trainer = CTCTrainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=processor.feature_extractor,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
# save the feature_extractor and the tokenizer
if is_main_process(training_args.local_rank):
processor.save_pretrained(training_args.output_dir)
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, MultiLingAdapterArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args, adapter_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1])
)
else:
model_args, data_args, training_args, adapter_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.train_file is not None or data_args.validation_file is not None:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
else:
# Downloading and loading the swag dataset from the hub.
datasets = load_dataset("swag", "regular")
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForMultipleChoice.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Setup adapters
if adapter_args.train_adapter:
task_name = "swag"
# check if adapter already exists otherwise add it
if task_name not in model.config.adapters:
# resolve adapter config
adapter_config = AdapterConfig.load(
adapter_args.adapter_config,
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
)
# load adapter from hub if specified
if adapter_args.load_adapter:
model.load_adapter(adapter_args.load_adapter, config=adapter_config, load_as=task_name)
else:
model.add_adapter(task_name, config=adapter_config)
# optionally load a pretrained language adapter
if adapter_args.load_lang_adapter:
# resolve language adapter config
lang_adapter_config = AdapterConfig.load(
adapter_args.lang_adapter_config,
non_linearity=adapter_args.lang_adapter_non_linearity,
reduction_factor=adapter_args.lang_adapter_reduction_factor,
)
# load language adapter from Hub
lang_adapter_name = model.load_adapter(
adapter_args.load_lang_adapter,
config=lang_adapter_config,
load_as=adapter_args.language,
)
else:
lang_adapter_name = None
# Freeze all model weights except of those in this adapter
model.train_adapter(task_name)
# Set the adapters to be used in every forward pass
if lang_adapter_name:
model.set_active_adapters(Fuse(lang_adapter_name, task_name))
else:
model.set_active_adapters(task_name)
else:
if adapter_args.load_adapter or adapter_args.load_lang_adapter:
raise ValueError(
"Adapters can only be loaded in adapters training mode."
"Use --train_adapter to enable adapter_training"
)
# When using your own dataset or a different dataset from swag, you will probably need to change this.
ending_names = [f"ending{i}" for i in range(4)]
context_name = "sent1"
question_header_name = "sent2"
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warn(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warn(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Preprocessing the datasets.
def preprocess_function(examples):
first_sentences = [[context] * 4 for context in examples[context_name]]
question_headers = examples[question_header_name]
second_sentences = [
[f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers)
]
# Flatten out
first_sentences = sum(first_sentences, [])
second_sentences = sum(second_sentences, [])
# Tokenize
tokenized_examples = tokenizer(
first_sentences,
second_sentences,
truncation=True,
max_length=max_seq_length,
padding="max_length" if data_args.pad_to_max_length else False,
)
# Un-flatten
return {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = (
default_data_collator
if data_args.pad_to_max_length
else DataCollatorForMultipleChoice(tokenizer=tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
)
# Metric
def compute_metrics(eval_predictions):
predictions, label_ids = eval_predictions
preds = np.argmax(predictions, axis=1)
return {"accuracy": (preds == label_ids).astype(np.float32).mean().item()}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
do_save_full_model=not adapter_args.train_adapter,
do_save_adapters=adapter_args.train_adapter,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name,
**tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, **tokenizer_kwargs)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if model_args.model_name_or_path:
model = AutoModelForMaskedLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
total_param_num = 0
for name, param in model.named_parameters():
if param.requires_grad:
param_num = param.numel()
total_param_num += param_num
# print(name, "\t", param_num)
print("total parameters num", total_param_num)
print("--" * 10)
freeze_layers = "0, 1, 2, 3, 4"
if freeze_layers is not "":
layer_indexes = [int(x) for x in freeze_layers.split(", ")]
total_freeze_param_num = 0
for param in model.distilbert.embeddings.parameters():
param.requires_grad = False
param_num = param.numel()
total_freeze_param_num += param_num
for layer_idx in layer_indexes:
for name, param in list(model.distilbert.transformer.
layer[layer_idx].named_parameters()):
param.requires_grad = False
param_num = param.numel()
total_freeze_param_num += param_num
# print(name, "\t", param_num)
print("froze layer", layer_idx)
print("total freeze parameters num", total_freeze_param_num)
print("left parameters num for training",
total_param_num - total_freeze_param_num)
# exit()
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples["text"] = [
line for line in examples["text"]
if len(line) > 0 and not line.isspace()
]
return tokenizer(examples["text"],
padding=padding,
truncation=True,
max_length=data_args.max_seq_length)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
# Add the chinese references if provided
if data_args.train_ref_file is not None:
tokenized_datasets["train"] = add_chinese_references(
tokenized_datasets["train"], data_args.train_ref_file)
if data_args.validation_ref_file is not None:
tokenized_datasets["validation"] = add_chinese_references(
tokenized_datasets["validation"], data_args.validation_ref_file)
# If we have ref files, need to avoid it removed by trainer
has_ref = data_args.train_ref_file or data_args.validation_ref_file
if has_ref:
training_args.remove_unused_columns = False
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForWholeWordMask(
tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"]
if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif model_args.model_name_or_path is not None and os.path.isdir(
model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_mlm_wwm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files in the summarization task, this script will use the first column for the full texts and the
# second column for the summaries (unless you specify column names for this with the `text_column` and
# `summary_column` arguments).
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (data_args.target_lang is not None and data_args.source_lang
is not None), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if data_args.task.startswith("translation") or isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if data_args.source_lang is not None:
tokenizer.src_lang = data_args.source_lang
if data_args.target_lang is not None:
tokenizer.tgt_lang = data_args.target_lang
# To serialize preprocess_function below, each of those four variables needs to be defined (even if we won't use
# them all).
source_lang, target_lang, text_column, summary_column = None, None, None, None
if data_args.task.startswith("summarization"):
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(
data_args.dataset_name, None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
else:
# Get the language codes for input/target.
lang_search = re.match("translation_([a-z]+)_to_([a-z]+)",
data_args.task)
if data_args.source_lang is not None:
source_lang = data_args.source_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a source language via --source_lang or rename your task 'translation_xx_to_yy'."
source_lang = lang_search.groups()[0]
if data_args.target_lang is not None:
target_lang = data_args.target_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a target language via --target_lang or rename your task 'translation_xx_to_yy'."
target_lang = lang_search.groups()[1]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warn(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
if data_args.task.startswith("translation"):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
else:
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric_name = "rouge" if data_args.task.startswith(
"summarization") else "sacrebleu"
metric = load_metric(metric_name)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
if metric_name == "rouge":
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
else: # sacrebleu
labels = [[label] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
if metric_name == "rouge":
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
else:
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Test ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(
test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(training_args.output_dir,
"test_generations.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
return results
def main(train_function):
# ----- Parse local_rank for torch.distributed.launch -----------
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int)
local_rank = parser.parse_args().local_rank
if local_rank is None:
local_rank = 0
# ----- Setup logging -----------
logger = logging.getLogger(__name__)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(local_rank) else logging.WARN)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# ----- Configurable Params -----------
# List of dicts with configuration for each dataset to be loaded
# see available datasets in the Hub: https://huggingface.co/datasets. sizes
# are of generated dataset, can be an order of magnitude larger after tokenization.
# Not all datasets can be concatenated without preprocessing, features must align
datasets_args = [
dict(path="wikitext", name="wikitext-2-raw-v1"), # 12.91 MB
# dict(path="wikitext", name="wikitext-103-raw-v1"), # 524 MB
# dict(path="ptb_text_only"), # 5.7 MB
# dict(path="bookcorpus"), # 4.63 GB
# dict(path="wikipedia"), # 35.38 GB
]
# Training params
# note: in V100 bs=8 uses 11/16 of available gpu mem, bs=12 uses 15/16
output_dir = os.path.expanduser("~/nta/results/bert")
training_args = TrainingArguments(
# Logging
output_dir=output_dir,
logging_first_step=True,
logging_steps=10, # also define eval_steps
eval_steps=10,
max_steps=30, # num_train_epochs replaced by steps
disable_tqdm=True,
run_name="debug_run", # used for wandb, not for Ray
# hyperparams
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
learning_rate=1e-4,
lr_scheduler_type="linear",
warmup_steps=500,
weight_decay=1e-6,
)
# Evaluate refers to evaluating perplexity on trained model in the validation set
# doesn't refer to finetuning and evaluating on downstream tasks such as GLUE
seed = random.randint(0, 1000000)
# Changing the tokenizer will result in re-tokenizing the dataset.
# As a reference, BERT tokenization will take ~ 3 hours for a 5GB dataset
config_class = BertConfig
tokenizer_name = "bert-base-cased"
# ----- Seed -----------
set_seed(seed)
print(f"Seed to reproduce: {seed}")
# ----- Dataset -----------
# Load multiple datasets and concatenate.
# using only 'train' and 'validation' sets, could also include 'test'
# if no split is defined, load_dataset returns DatasetDict with all available splits
train_datasets = [load_dataset(**args, split="train") for args in datasets_args]
val_datasets = [load_dataset(**args, split="validation") for args in datasets_args]
dataset = DatasetDict()
dataset["train"] = concatenate_datasets(train_datasets)
dataset["validation"] = concatenate_datasets(val_datasets)
def load_and_split_dataset(dataset_args, split_percentage=5):
"""Alternative: if no validation set available, manuallly split the train set"""
dataset = DatasetDict()
dataset["train"] = load_dataset(
**dataset_args, split=f"train[{split_percentage}%:]"
)
dataset["validation"] = load_dataset(
**dataset_args, split=f"train[:{split_percentage}%]"
)
return dataset
# ----- Load Model -----------
# Load model
config = config_class()
model = AutoModelForMaskedLM.from_config(config)
# Load tokenizer
# use_fast falls back to tokenizer lib implementation under the hood
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, use_fast=True)
model.resize_token_embeddings(len(tokenizer))
# ----- Preprocess dataset -----------
# Only use the text column name when doing language modeling
# this feature might have a different name depending on the dataset
# might need to change column names prior to concatenating, if that is the case
column_names = dataset["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# Setting overwrite_cache to False will retokenize the dataset.
# do not overwrite cache if using shared cache repository.
overwrite_cache = False
preprocessing_num_workers = None
# We tokenize every text, then concatenate them together before splitting in smaller
# parts. We use `return_special_tokens_mask=True` given
# DataCollatorForLanguageModeling is more efficient when it
# receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
tokenized_dataset = dataset.map(
tokenize_function,
batched=True,
remove_columns=column_names,
num_proc=preprocessing_num_workers,
load_from_cache_file=not overwrite_cache,
)
# Main data processing function that will concatenate all texts from our dataset and
# generate chunks of max_seq_length.
max_seq_length = tokenizer.model_max_length
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it
# instead of this drop, you can customize this part to your needs.
total_length = (total_length // max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [t[i : i + max_seq_length] for i in
range(0, total_length, max_seq_length)]
for k, t in concatenated_examples.items()
}
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so
# group_texts throws away a remainder for each of those groups of 1,000 texts.
# You can adjust batch_size here but a higher value will be slower to preprocess.
tokenized_dataset = tokenized_dataset.map(
group_texts,
batched=True,
num_proc=preprocessing_num_workers,
load_from_cache_file=not overwrite_cache,
)
# Data collator
# This one will take care of randomly masking the tokens.
# Q: what about dynamic masking, used in Roberta?
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm_probability=0.15
)
# ----- Setup Trainer -----------
# Initialize Trainer. Similar to Vernon's Experiment class.
# dataloader and training loop are contained in Trainer abstraction
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_dataset["train"],
eval_dataset=tokenized_dataset["validation"],
tokenizer=tokenizer,
data_collator=data_collator,
)
# ----- Functions to train and evaluate -----------
if train_function == "huggingface":
# Tested
run_hf(trainer, logger, output_dir, save_model=True, evaluate=True)
elif train_function == "ray_single_node":
# Tested
run_ray_single_instance(
trainer,
logger,
name="bert_test",
config=None,
num_samples=1,
local_dir=os.path.expanduser("~/nta/results/experiments/transformers"),
keep_checkpoints_num=1,
resources_per_trial={"cpu": 8},
# note: checkpoint arguments cannot be used with a checkpointable function
)
elif train_function == "ray_multiple_nodes":
# Untested
run_ray_distributed(
trainer,
logger,
name="bert_test",
config=None,
num_samples=1,
local_dir=os.path.expanduser("~/nta/results/experiments/transformers"),
keep_checkpoints_num=1,
queue_trials=True,
verbose=2,
resources_per_trial={"gpu": 4},
)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
#
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.task_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset("glue",
data_args.task_name,
cache_dir=model_args.cache_dir)
else:
# Loading a dataset from your local files.
# CSV/JSON training and evaluation files are needed.
data_files = {
"train": data_args.train_file,
"validation": data_args.validation_file
}
# Get the test dataset: you can provide your own CSV/JSON test file (see below)
# when you use `do_predict` without specifying a GLUE benchmark task.
if training_args.do_predict:
if data_args.test_file is not None:
train_extension = data_args.train_file.split(".")[-1]
test_extension = data_args.test_file.split(".")[-1]
assert (
test_extension == train_extension
), "`test_file` should have the same extension (csv or json) as `train_file`."
data_files["test"] = data_args.test_file
else:
raise ValueError(
"Need either a GLUE task or a test file for `do_predict`.")
for key in data_files.keys():
logger.info(f"load a local file for {key}: {data_files[key]}")
if data_args.train_file.endswith(".csv"):
# Loading a dataset from local csv files
datasets = load_dataset("csv",
data_files=data_files,
cache_dir=model_args.cache_dir)
else:
# Loading a dataset from local json files
datasets = load_dataset("json",
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if data_args.task_name is not None:
is_regression = data_args.task_name == "stsb"
if not is_regression:
label_list = datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = datasets["train"].features["label"].dtype in [
"float32", "float64"
]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Preprocessing the datasets
if data_args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[data_args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [
name for name in datasets["train"].column_names if name != "label"
]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Padding strategy
if data_args.pad_to_max_length:
padding = "max_length"
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (model.config.label2id !=
PretrainedConfig(num_labels=num_labels).label2id
and data_args.task_name is not None and not is_regression):
# Some have all caps in their config, some don't.
label_name_to_id = {
k.lower(): v
for k, v in model.config.label2id.items()
}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
label_to_id = {
i: int(label_name_to_id[label_list[i]])
for i in range(num_labels)
}
else:
logger.warning(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
elif data_args.task_name is None and not is_regression:
label_to_id = {v: i for i, v in enumerate(label_list)}
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_function(examples):
# Tokenize the texts
args = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*args,
padding=padding,
max_length=max_seq_length,
truncation=True)
# Map labels to IDs (not necessary for GLUE tasks)
if label_to_id is not None and "label" in examples:
result["label"] = [(label_to_id[l] if l != -1 else -1)
for l in examples["label"]]
return result
datasets = datasets.map(preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache)
if training_args.do_train:
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in datasets and "validation_matched" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation_matched" if data_args.task_name ==
"mnli" else "validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
if training_args.do_predict or data_args.task_name is not None or data_args.test_file is not None:
if "test" not in datasets and "test_matched" not in datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = datasets["test_matched" if data_args.task_name ==
"mnli" else "test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
# Log a few random samples from the training set:
if training_args.do_train:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# Get the metric function
if data_args.task_name is not None:
metric = load_metric("glue", data_args.task_name)
# TODO: When datasets metrics include regular accuracy, make an else here and remove special branch from
# compute_metrics
# You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
# predictions and label_ids field) and has to return a dictionary string to float.
def compute_metrics(p: EvalPrediction):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
preds = np.squeeze(preds) if is_regression else np.argmax(preds,
axis=1)
if data_args.task_name is not None:
result = metric.compute(predictions=preds, references=p.label_ids)
if len(result) > 1:
result["combined_score"] = np.mean(list(
result.values())).item()
return result
elif is_regression:
return {"mse": ((preds - p.label_ids)**2).mean().item()}
else:
return {
"accuracy":
(preds == p.label_ids).astype(np.float32).mean().item()
}
# Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
if data_args.pad_to_max_length:
data_collator = default_data_collator
elif training_args.fp16:
data_collator = DataCollatorWithPadding(tokenizer,
pad_to_multiple_of=8)
else:
data_collator = None
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.save_model() # Saves the tokenizer too for easy upload
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = [data_args.task_name]
eval_datasets = [eval_dataset]
if data_args.task_name == "mnli":
tasks.append("mnli-mm")
eval_datasets.append(datasets["validation_mismatched"])
for eval_dataset, task in zip(eval_datasets, tasks):
metrics = trainer.evaluate(eval_dataset=eval_dataset)
max_eval_samples = (data_args.max_eval_samples
if data_args.max_eval_samples is not None else
len(eval_dataset))
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = [data_args.task_name]
predict_datasets = [predict_dataset]
if data_args.task_name == "mnli":
tasks.append("mnli-mm")
predict_datasets.append(datasets["test_mismatched"])
for predict_dataset, task in zip(predict_datasets, tasks):
# Removing the `label` columns because it contains -1 and Trainer won't like that.
predict_dataset.remove_columns_("label")
predictions = trainer.predict(
predict_dataset, metric_key_prefix="predict").predictions
predictions = np.squeeze(
predictions) if is_regression else np.argmax(predictions,
axis=1)
output_predict_file = os.path.join(training_args.output_dir,
f"predict_results_{task}.txt")
if trainer.is_world_process_zero():
with open(output_predict_file, "w") as writer:
logger.info(f"***** Predict results {task} *****")
writer.write("index\tprediction\n")
for index, item in enumerate(predictions):
if is_regression:
writer.write(f"{index}\t{item:3.3f}\n")
else:
item = label_list[item]
writer.write(f"{index}\t{item}\n")
if training_args.push_to_hub:
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tags": "text-classification"
}
if data_args.task_name is not None:
kwargs["language"] = "en"
kwargs["dataset_tags"] = "glue"
kwargs["dataset_args"] = data_args.task_name
kwargs["dataset"] = f"GLUE {data_args.task_name.upper()}"
trainer.push_to_hub(**kwargs)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 1. First, let's load the dataset
raw_datasets = IterableDatasetDict()
raw_column_names = {}
def load_streaming_dataset(split, sampling_rate, **kwargs):
if "+" in split:
dataset_splits = [
load_dataset(split=split_name, **kwargs)
for split_name in split.split("+")
]
# `features` and `cast_column` won't be available after interleaving, so we'll use them here
features = dataset_splits[0].features
# make sure that the dataset decodes audio with a correct sampling rate
dataset_splits = [
dataset.cast_column(
data_args.audio_column_name,
datasets.features.Audio(sampling_rate=sampling_rate))
for dataset in dataset_splits
]
interleaved_dataset = interleave_datasets(dataset_splits)
return interleaved_dataset, features
else:
dataset = load_dataset(split=split, **kwargs)
features = dataset.features
# make sure that the dataset decodes audio with a correct sampling rate
dataset = dataset.cast_column(
data_args.audio_column_name,
datasets.features.Audio(sampling_rate=sampling_rate))
return dataset, features
# `datasets` takes care of automatically loading and resampling the audio,
# so we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
if training_args.do_train:
raw_datasets["train"], train_features = load_streaming_dataset(
path=data_args.dataset_name,
name=data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
streaming=True,
sampling_rate=feature_extractor.sampling_rate,
)
raw_column_names["train"] = list(train_features.keys())
if data_args.audio_column_name not in raw_column_names["train"]:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'."
" Make sure to set `--audio_column_name` to the correct audio column - one of"
f" {', '.join(raw_column_names['train'])}.")
if data_args.text_column_name not in raw_column_names["train"]:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_column_names['train'])}.")
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].take(
range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"], eval_features = load_streaming_dataset(
path=data_args.dataset_name,
name=data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
streaming=True,
sampling_rate=feature_extractor.sampling_rate,
)
raw_column_names["eval"] = list(eval_features.keys())
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].take(
range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
text_column_name = data_args.text_column_name
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[text_column_name]).lower() + " "
else:
batch["target_text"] = batch[text_column_name].lower() + " "
return batch
with training_args.main_process_first(
desc="dataset map special characters removal"):
for split, dataset in raw_datasets.items():
raw_datasets[split] = dataset.map(
remove_special_characters, ).remove_columns([text_column_name])
# 3. Next, let's load the config as we might need it to create
# the tokenizer
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# 4. Now we can instantiate the tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
tokenizer_name_or_path = model_args.tokenizer_name_or_path
if tokenizer_name_or_path is None:
raise ValueError(
"Tokenizer has to be created before training in streaming mode. Please specify --tokenizer_name_or_path"
)
# load feature_extractor and tokenizer
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
config=config,
use_auth_token=data_args.use_auth_token,
)
# adapt config
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": tokenizer.pad_token_id,
"vocab_size": len(tokenizer),
"activation_dropout": model_args.activation_dropout,
})
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 5. Now we preprocess the datasets including loading the audio, resampling and normalization
audio_column_name = data_args.audio_column_name
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"],
**additional_kwargs).input_ids
return batch
vectorized_datasets = IterableDatasetDict()
with training_args.main_process_first(desc="dataset map preprocessing"):
for split, dataset in raw_datasets.items():
vectorized_datasets[split] = (
dataset.map(prepare_dataset).remove_columns(
raw_column_names[split] +
["target_text"]).with_format("torch"))
if split == "train":
vectorized_datasets[split] = vectorized_datasets[
split].shuffle(
buffer_size=data_args.shuffle_buffer_size,
seed=training_args.seed,
)
# 6. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {
metric: load_metric(metric)
for metric in data_args.eval_metrics
}
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {
k: v.compute(predictions=pred_str, references=label_str)
for k, v in eval_metrics.items()
}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
data_collator = DataCollatorCTCWithPadding(processor=processor,
max_length=max_input_length)
# trainer callback to reinitialize and reshuffle the streamable datasets at the beginning of each epoch
class ShuffleCallback(TrainerCallback):
def on_epoch_begin(self, args, state, control, train_dataloader,
**kwargs):
if isinstance(train_dataloader.dataset, IterableDatasetShard):
pass # set_epoch() is handled by the Trainer
elif isinstance(train_dataloader.dataset, IterableDataset):
train_dataloader.dataset.set_epoch(
train_dataloader.dataset._epoch + 1)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=processor,
callbacks=[ShuffleCallback()],
)
# 7. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
if data_args.max_train_samples:
metrics["train_samples"] = data_args.max_train_samples
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
if data_args.max_eval_samples:
metrics["eval_samples"] = data_args.max_eval_samples
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
kwargs = {
"finetuned_from":
model_args.model_name_or_path,
"tasks":
"speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args":
(f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
f" {data_args.eval_split_name}"),
"dataset":
f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = config_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
